{
 "generated_at": "2026-10-21T11:54:02Z",
 "vendor_efficiency": {
  "Apple": 0.831,
  "all": 0.688,
  "NVIDIA": 0.656,
  "AMD": 0.722,
  "Intel": 0.55
 },
 "hardware": {
  "apple-m1-7c": {
   "slug": "apple-m1-7c",
   "name": "Apple M1 (7-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 68.25,
   "chip": "M1",
   "gpu_cores": 7,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": null,
   "eff_bw": 56.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 7.92,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m1-8c": {
   "slug": "apple-m1-8c",
   "name": "Apple M1 (8-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 68.25,
   "chip": "M1",
   "gpu_cores": 8,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": null,
   "eff_bw": 56.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 7.91,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m1-pro-14c": {
   "slug": "apple-m1-pro-14c",
   "name": "Apple M1 Pro (14-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M1 Pro",
   "gpu_cores": 14,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 204.8",
   "eff_bw": 171.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 12.75,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "apple-m1-pro-16c": {
   "slug": "apple-m1-pro-16c",
   "name": "Apple M1 Pro (16-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M1 Pro",
   "gpu_cores": 16,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 204.8",
   "eff_bw": 171.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 12.75,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "apple-m1-max-24c": {
   "slug": "apple-m1-max-24c",
   "name": "Apple M1 Max (24-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M1 Max",
   "gpu_cores": 24,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 409.6",
   "eff_bw": 303.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 22.55,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "apple-m1-max-32c": {
   "slug": "apple-m1-max-32c",
   "name": "Apple M1 Max (32-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M1 Max",
   "gpu_cores": 32,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 409.6",
   "eff_bw": 310.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 23.03,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "apple-m1-ultra-48c": {
   "slug": "apple-m1-ultra-48c",
   "name": "Apple M1 Ultra (48-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 800,
   "chip": "M1 Ultra",
   "gpu_cores": 48,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 819.2",
   "eff_bw": 457.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 33.92,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "apple-m1-ultra-64c": {
   "slug": "apple-m1-ultra-64c",
   "name": "Apple M1 Ultra (64-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 800,
   "chip": "M1 Ultra",
   "gpu_cores": 64,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 819.2",
   "eff_bw": 498.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 37.01,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "apple-m2-8c": {
   "slug": "apple-m2-8c",
   "name": "Apple M2 (8-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M2",
   "gpu_cores": 8,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 102.4",
   "eff_bw": 87.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 12.18,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m2-10c": {
   "slug": "apple-m2-10c",
   "name": "Apple M2 (10-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M2",
   "gpu_cores": 10,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 102.4",
   "eff_bw": 90.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 6.72,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "apple-m2-pro-16c": {
   "slug": "apple-m2-pro-16c",
   "name": "Apple M2 Pro (16-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M2 Pro",
   "gpu_cores": 16,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 204.8",
   "eff_bw": 168.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 12.47,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "apple-m2-pro-19c": {
   "slug": "apple-m2-pro-19c",
   "name": "Apple M2 Pro (19-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M2 Pro",
   "gpu_cores": 19,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 204.8",
   "eff_bw": 176.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 13.06,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m2-max-30c": {
   "slug": "apple-m2-max-30c",
   "name": "Apple M2 Max (30-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M2 Max",
   "gpu_cores": 30,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 409.6",
   "eff_bw": 325.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 24.16,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "apple-m2-max-38c": {
   "slug": "apple-m2-max-38c",
   "name": "Apple M2 Max (38-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M2 Max",
   "gpu_cores": 38,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 409.6",
   "eff_bw": 332.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 24.65,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "apple-m2-ultra-60c": {
   "slug": "apple-m2-ultra-60c",
   "name": "Apple M2 Ultra (60-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 800,
   "chip": "M2 Ultra",
   "gpu_cores": 60,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 819.2",
   "eff_bw": 537.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 39.86,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "apple-m2-ultra-76c": {
   "slug": "apple-m2-ultra-76c",
   "name": "Apple M2 Ultra (76-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 800,
   "chip": "M2 Ultra",
   "gpu_cores": 76,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 819.2",
   "eff_bw": 552.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 41.02,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m3-10c": {
   "slug": "apple-m3-10c",
   "name": "Apple M3 (10-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M3",
   "gpu_cores": 10,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "theoretical 102.4",
   "eff_bw": 87.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 12.27,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m3-pro-14c": {
   "slug": "apple-m3-pro-14c",
   "name": "Apple M3 Pro (14-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 150,
   "chip": "M3 Pro",
   "gpu_cores": 14,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "theoretical 153.6",
   "eff_bw": 124.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 17.44,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "apple-m3-pro-18c": {
   "slug": "apple-m3-pro-18c",
   "name": "Apple M3 Pro (18-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 150,
   "chip": "M3 Pro",
   "gpu_cores": 18,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "theoretical 153.6",
   "eff_bw": 133.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 9.89,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "apple-m3-max-30c": {
   "slug": "apple-m3-max-30c",
   "name": "Apple M3 Max (30-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 300,
   "chip": "M3 Max",
   "gpu_cores": 30,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "30-core GPU bin; theoretical 307.2",
   "eff_bw": 263.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 19.54,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "apple-m3-max-40c": {
   "slug": "apple-m3-max-40c",
   "name": "Apple M3 Max (40-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M3 Max",
   "gpu_cores": 40,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "40-core GPU bin; theoretical 409.6",
   "eff_bw": 338.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 25.09,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "apple-m3-ultra-60c": {
   "slug": "apple-m3-ultra-60c",
   "name": "Apple M3 Ultra (60-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 819,
   "chip": "M3 Ultra",
   "gpu_cores": 60,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "Wikipedia 819.3; corrected from 800 (llama.cpp #4167 table)",
   "eff_bw": 569.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 42.24,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "apple-m3-ultra-80c": {
   "slug": "apple-m3-ultra-80c",
   "name": "Apple M3 Ultra (80-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 819,
   "chip": "M3 Ultra",
   "gpu_cores": 80,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "Wikipedia 819.3; corrected from 800 (llama.cpp #4167 table)",
   "eff_bw": 536.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 39.78,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "apple-m4-10c": {
   "slug": "apple-m4-10c",
   "name": "Apple M4 (10-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 120,
   "chip": "M4",
   "gpu_cores": 10,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": null,
   "eff_bw": 100.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 7.43,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 12
  },
  "apple-m4-pro-16c": {
   "slug": "apple-m4-pro-16c",
   "name": "Apple M4 Pro (16-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 273,
   "chip": "M4 Pro",
   "gpu_cores": 16,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": null,
   "eff_bw": 231.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 17.19,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "apple-m4-pro-20c": {
   "slug": "apple-m4-pro-20c",
   "name": "Apple M4 Pro (20-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 273,
   "chip": "M4 Pro",
   "gpu_cores": 20,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": null,
   "eff_bw": 231.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 17.18,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 12
  },
  "apple-m4-max-32c": {
   "slug": "apple-m4-max-32c",
   "name": "Apple M4 Max (32-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 410,
   "chip": "M4 Max",
   "gpu_cores": 32,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": "32-core GPU bin",
   "eff_bw": 327.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 24.29,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "apple-m4-max-40c": {
   "slug": "apple-m4-max-40c",
   "name": "Apple M4 Max (40-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 546,
   "chip": "M4 Max",
   "gpu_cores": 40,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": "40-core GPU bin",
   "eff_bw": 426.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 31.64,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 12
  },
  "apple-m5-10c": {
   "slug": "apple-m5-10c",
   "name": "Apple M5 (10-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 153.6,
   "chip": "M5",
   "gpu_cores": 10,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": null,
   "eff_bw": 131.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 18.42,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "apple-m5-pro-16c": {
   "slug": "apple-m5-pro-16c",
   "name": "Apple M5 Pro (16-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 307,
   "chip": "M5 Pro",
   "gpu_cores": 16,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": "up to",
   "eff_bw": 288.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 21.39,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "apple-m5-pro-20c": {
   "slug": "apple-m5-pro-20c",
   "name": "Apple M5 Pro (20-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 307,
   "chip": "M5 Pro",
   "gpu_cores": 20,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": "up to",
   "eff_bw": 290.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 21.55,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "apple-m5-max-40c": {
   "slug": "apple-m5-max-40c",
   "name": "Apple M5 Max (40-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 614,
   "chip": "M5 Max",
   "gpu_cores": 40,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": "40-core GPU bin",
   "eff_bw": 518.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 72.42,
    "size_gb": 7.16,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "apple-m5-ultra-80c": {
   "slug": "apple-m5-ultra-80c",
   "name": "Apple M5 Ultra (80-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 1228.8,
   "chip": "M5 Ultra",
   "gpu_cores": 80,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": "DERIVED: 2x M5 Max (Wikipedia lists LPDDR5X-9600; no explicit Ultra bandwidth cell); matches llama.cpp #4167 table (1228)",
   "eff_bw": 991.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 73.6,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "apple-m6-8c": {
   "slug": "apple-m6-8c",
   "name": "Apple M6 (8-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 170,
   "chip": "M6",
   "gpu_cores": 8,
   "spec_source": "Wikipedia (Apple M6)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M6",
   "spec_note": "Apple-rated peak per Wikipedia",
   "eff_bw": 156.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "F16",
    "backend": "Metal",
    "tg": 11.59,
    "size_gb": 13.48,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/4167",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-rtx-5090": {
   "slug": "nvidia-rtx-5090",
   "name": "NVIDIA RTX 5090",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 32,
   "bandwidth_gbs": 1792,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 1148.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 300.4,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "nvidia-rtx-pro-6000-blackwell": {
   "slug": "nvidia-rtx-pro-6000-blackwell",
   "name": "NVIDIA RTX PRO 6000 Blackwell",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 96,
   "bandwidth_gbs": 1792,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 1074.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 281.11,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-h100-sxm": {
   "slug": "nvidia-h100-sxm",
   "name": "NVIDIA H100 SXM",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 80,
   "bandwidth_gbs": 3350,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia 3352",
   "eff_bw": 1073.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 280.74,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-a100-80gb-pcie": {
   "slug": "nvidia-a100-80gb-pcie",
   "name": "NVIDIA A100 80GB PCIe",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 80,
   "bandwidth_gbs": 1935,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/data-center/a100/",
   "spec_note": "NVIDIA A100 spec table (80GB PCIe 1,935 GB/s)",
   "eff_bw": 767.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 200.9,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-4090-d": {
   "slug": "nvidia-rtx-4090-d",
   "name": "NVIDIA RTX 4090 D",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 1008,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 732.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 191.57,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-rtx-4090": {
   "slug": "nvidia-rtx-4090",
   "name": "NVIDIA RTX 4090",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 1008,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 726.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 190.1,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 14
  },
  "nvidia-rtx-5080": {
   "slug": "nvidia-rtx-5080",
   "name": "NVIDIA RTX 5080",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 960,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 734.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 192.17,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 15
  },
  "nvidia-rtx-5070-ti": {
   "slug": "nvidia-rtx-5070-ti",
   "name": "NVIDIA RTX 5070 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 896,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 697.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 182.43,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "nvidia-rtx-6000-ada": {
   "slug": "nvidia-rtx-6000-ada",
   "name": "NVIDIA RTX 6000 Ada",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 48,
   "bandwidth_gbs": 960,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 686.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 179.47,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "nvidia-rtx-3090-ti": {
   "slug": "nvidia-rtx-3090-ti",
   "name": "NVIDIA RTX 3090 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 1008,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 658.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 172.26,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-3090": {
   "slug": "nvidia-rtx-3090",
   "name": "NVIDIA RTX 3090",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 936,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 656.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 171.61,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "nvidia-rtx-5090-laptop": {
   "slug": "nvidia-rtx-5090-laptop",
   "name": "NVIDIA RTX 5090 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 24,
   "bandwidth_gbs": 896,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/geforce/laptops/50-series/",
   "spec_note": "NVIDIA spec table",
   "eff_bw": 604.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 158.14,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 9
  },
  "nvidia-l40": {
   "slug": "nvidia-l40",
   "name": "NVIDIA L40",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 48,
   "bandwidth_gbs": 864,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 587.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 153.76,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-4080-super": {
   "slug": "nvidia-rtx-4080-super",
   "name": "NVIDIA RTX 4080 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 736,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 574.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 150.2,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "nvidia-rtx-4080": {
   "slug": "nvidia-rtx-4080",
   "name": "NVIDIA RTX 4080",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 717,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 548.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 143.47,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-3080": {
   "slug": "nvidia-rtx-3080",
   "name": "NVIDIA RTX 3080",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 10,
   "bandwidth_gbs": 760,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 557.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 145.74,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "nvidia-rtx-a6000": {
   "slug": "nvidia-rtx-a6000",
   "name": "NVIDIA RTX A6000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 48,
   "bandwidth_gbs": 768,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 553.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 144.87,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-rtx-4070-ti-super": {
   "slug": "nvidia-rtx-4070-ti-super",
   "name": "NVIDIA RTX 4070 Ti SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 519.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 135.81,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-rtx-pro-4000-blackwell": {
   "slug": "nvidia-rtx-pro-4000-blackwell",
   "name": "NVIDIA RTX PRO 4000 Blackwell",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 24,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 521.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 136.38,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-a5000": {
   "slug": "nvidia-rtx-a5000",
   "name": "NVIDIA RTX A5000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 24,
   "bandwidth_gbs": 768,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 536.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 140.43,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-tesla-v100-32gb": {
   "slug": "nvidia-tesla-v100-32gb",
   "name": "NVIDIA Tesla V100 32GB",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 32,
   "bandwidth_gbs": 900,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 543.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 142.13,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "nvidia-rtx-5070": {
   "slug": "nvidia-rtx-5070",
   "name": "NVIDIA RTX 5070",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 490.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 128.21,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-a40": {
   "slug": "nvidia-a40",
   "name": "NVIDIA A40",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 48,
   "bandwidth_gbs": 696,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "695.8",
   "eff_bw": 482.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 126.24,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-a30": {
   "slug": "nvidia-a30",
   "name": "NVIDIA A30",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 24,
   "bandwidth_gbs": 933,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "933.1",
   "eff_bw": 504.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 131.93,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-titan-v": {
   "slug": "nvidia-titan-v",
   "name": "NVIDIA TITAN V",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 653,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 428.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 112.17,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-rtx-2080-ti": {
   "slug": "nvidia-rtx-2080-ti",
   "name": "NVIDIA RTX 2080 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 11,
   "bandwidth_gbs": 616,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 417.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 109.17,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-quadro-rtx-6000": {
   "slug": "nvidia-quadro-rtx-6000",
   "name": "NVIDIA Quadro RTX 6000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 24,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 399.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 104.38,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-quadro-rtx-8000": {
   "slug": "nvidia-quadro-rtx-8000",
   "name": "NVIDIA Quadro RTX 8000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 48,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 396.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 103.63,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-rtx-a4500": {
   "slug": "nvidia-rtx-a4500",
   "name": "NVIDIA RTX A4500",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 20,
   "bandwidth_gbs": 640,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 393.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 103.0,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-5060-ti": {
   "slug": "nvidia-rtx-5060-ti",
   "name": "NVIDIA RTX 5060 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 370.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 97.01,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 11
  },
  "nvidia-rtx-2070-super": {
   "slug": "nvidia-rtx-2070-super",
   "name": "NVIDIA RTX 2070 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 336.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 88.06,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-rtx-a4000": {
   "slug": "nvidia-rtx-a4000",
   "name": "NVIDIA RTX A4000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 16,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 325.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 85.17,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "nvidia-titan-xp": {
   "slug": "nvidia-titan-xp",
   "name": "NVIDIA TITAN Xp",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 548,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 290.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 76.08,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-rtx-3060": {
   "slug": "nvidia-rtx-3060",
   "name": "NVIDIA RTX 3060",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 360,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "360 GB/s = 15 GT/s x 192-bit (12 GB model); Wikipedia's 12GB row bandwidth cell (336) conflicts with its own clock/bus cells",
   "eff_bw": 308.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 80.59,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "nvidia-quadro-rtx-4000": {
   "slug": "nvidia-quadro-rtx-4000",
   "name": "NVIDIA Quadro RTX 4000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 8,
   "bandwidth_gbs": 416,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 258.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 67.62,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-rtx-4060-ti": {
   "slug": "nvidia-rtx-4060-ti",
   "name": "NVIDIA RTX 4060 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 244.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 64.03,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 8
  },
  "nvidia-gtx-1080-ti": {
   "slug": "nvidia-gtx-1080-ti",
   "name": "NVIDIA GTX 1080 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 11,
   "bandwidth_gbs": 484,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 273.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 71.58,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-rtx-4000-ada": {
   "slug": "nvidia-rtx-4000-ada",
   "name": "NVIDIA RTX 4000 Ada",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 20,
   "bandwidth_gbs": 360,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 236.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 61.83,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-2060-super": {
   "slug": "nvidia-rtx-2060-super",
   "name": "NVIDIA RTX 2060 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 233.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 61.13,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-tesla-p100": {
   "slug": "nvidia-tesla-p100",
   "name": "NVIDIA Tesla P100",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 16,
   "bandwidth_gbs": 732,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "732 (16GB HBM2)",
   "eff_bw": 254.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 66.48,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "nvidia-dgx-spark-gb10": {
   "slug": "nvidia-dgx-spark-gb10",
   "name": "NVIDIA DGX Spark (GB10)",
   "vendor": "NVIDIA",
   "kind": "apu",
   "memory_gb": 128,
   "bandwidth_gbs": 273,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/",
   "spec_note": "NVIDIA spec table",
   "eff_bw": 238.3,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 Coder 7B",
    "quant": "Q8_0",
    "backend": "CUDA",
    "tg": 29.43,
    "size_gb": 8.1,
    "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md",
    "test": "tg32"
   },
   "n_rows": 23
  },
  "nvidia-tesla-p40": {
   "slug": "nvidia-tesla-p40",
   "name": "NVIDIA Tesla P40",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 24,
   "bandwidth_gbs": 346,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "corrected 347 -> 346 (Wikipedia 345.6)",
   "eff_bw": 226.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 59.36,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-rtx-2000-ada": {
   "slug": "nvidia-rtx-2000-ada",
   "name": "NVIDIA RTX 2000 Ada",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 16,
   "bandwidth_gbs": 224,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 193.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 50.71,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-tesla-t4": {
   "slug": "nvidia-tesla-t4",
   "name": "NVIDIA Tesla T4",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 16,
   "bandwidth_gbs": 320,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 177.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 46.38,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-rtx-4050-laptop": {
   "slug": "nvidia-rtx-4050-laptop",
   "name": "NVIDIA RTX 4050 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 6,
   "bandwidth_gbs": 192,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (168-192)",
   "eff_bw": 167.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 43.72,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-gtx-1660": {
   "slug": "nvidia-gtx-1660",
   "name": "NVIDIA GTX 1660",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": 192,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 158.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 41.43,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-tesla-m40": {
   "slug": "nvidia-tesla-m40",
   "name": "NVIDIA Tesla M40",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 24,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 159.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 41.68,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-gtx-1070-ti": {
   "slug": "nvidia-gtx-1070-ti",
   "name": "NVIDIA GTX 1070 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 166.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 43.42,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-jetson-agx-orin-64gb": {
   "slug": "nvidia-jetson-agx-orin-64gb",
   "name": "NVIDIA Jetson AGX Orin 64GB",
   "vendor": "NVIDIA",
   "kind": "apu",
   "memory_gb": 64,
   "bandwidth_gbs": 204.8,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/",
   "spec_note": "NVIDIA spec table 204.8",
   "eff_bw": 137.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 35.88,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-tesla-p4": {
   "slug": "nvidia-tesla-p4",
   "name": "NVIDIA Tesla P4",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 8,
   "bandwidth_gbs": 192,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 127.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 33.29,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-p106-100": {
   "slug": "nvidia-p106-100",
   "name": "NVIDIA P106-100",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 6,
   "bandwidth_gbs": 192,
   "spec_source": "cputronic.com",
   "spec_url": "https://cputronic.com/en/gpu/compare/nvidia-p106-100-vs-nvidia-p104-100",
   "spec_note": "third-party spec DB: 6 GB GDDR5, 192-bit, 192.2 GB/s (no manufacturer page found)",
   "eff_bw": 117.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 30.79,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-gtx-1060": {
   "slug": "nvidia-gtx-1060",
   "name": "NVIDIA GTX 1060",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": 192,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 107.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 28.18,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-t1000": {
   "slug": "nvidia-t1000",
   "name": "NVIDIA T1000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 4,
   "bandwidth_gbs": 160,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 106.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 27.82,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-quadro-p2000": {
   "slug": "nvidia-quadro-p2000",
   "name": "NVIDIA Quadro P2000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 5,
   "bandwidth_gbs": 140,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 90.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 23.77,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-quadro-p1000": {
   "slug": "nvidia-quadro-p1000",
   "name": "NVIDIA Quadro P1000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 4,
   "bandwidth_gbs": 80,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "corrected 82 -> 80 (Wikipedia 80.2)",
   "eff_bw": 53.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 13.99,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-tesla-k80": {
   "slug": "nvidia-tesla-k80",
   "name": "NVIDIA Tesla K80",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 12,
   "bandwidth_gbs": 240,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "per GPU (card has 2 GPUs)",
   "eff_bw": 54.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 14.27,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-rtx-pro-4500-blackwell": {
   "slug": "nvidia-rtx-pro-4500-blackwell",
   "name": "NVIDIA RTX PRO 4500 Blackwell",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 32,
   "bandwidth_gbs": 896,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 645.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "CUDA",
    "tg": 168.9,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15013",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-7900-xtx": {
   "slug": "amd-radeon-rx-7900-xtx",
   "name": "AMD Radeon RX 7900 XTX",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 960,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 729.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 190.92,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "nvidia-a100": {
   "slug": "nvidia-a100",
   "name": "Nvidia A100",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 652.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 170.56,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-ai-pro-r9700": {
   "slug": "amd-radeon-ai-pro-r9700",
   "name": "AMD Radeon AI PRO R9700",
   "vendor": "AMD",
   "kind": "workstation",
   "memory_gb": 32,
   "bandwidth_gbs": 640,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 583.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 152.7,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-rx-9070-xt": {
   "slug": "amd-radeon-rx-9070-xt",
   "name": "AMD Radeon RX 9070 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 640,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 524.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 137.11,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-rx-7900-xt": {
   "slug": "amd-radeon-rx-7900-xt",
   "name": "AMD Radeon RX 7900 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 20,
   "bandwidth_gbs": 800,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 470.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 123.18,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "amd-radeon-rx-9070": {
   "slug": "amd-radeon-rx-9070",
   "name": "AMD Radeon RX 9070",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 640,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 457.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 119.71,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-7800-xt": {
   "slug": "amd-radeon-rx-7800-xt",
   "name": "AMD Radeon RX 7800 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 624,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 477.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 124.86,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 10
  },
  "amd-radeon-rx-7900-gre": {
   "slug": "amd-radeon-rx-7900-gre",
   "name": "AMD Radeon RX 7900 GRE",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 576,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 443.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 116.11,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "apple-m3-ultra": {
   "slug": "apple-m3-ultra",
   "name": "Apple M3 Ultra",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 819,
   "chip": "M3 Ultra",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "Wikipedia 819.3; corrected from 800 (llama.cpp #4167 table)",
   "eff_bw": 441.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 115.54,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-arc-pro-b70": {
   "slug": "intel-arc-pro-b70",
   "name": "Intel Arc Pro B70",
   "vendor": "Intel",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 436.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 114.19,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-4070-ti": {
   "slug": "nvidia-rtx-4070-ti",
   "name": "NVIDIA RTX 4070 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 504,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 433.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 113.52,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-4070-super": {
   "slug": "nvidia-rtx-4070-super",
   "name": "NVIDIA RTX 4070 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 504,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 428.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 112.0,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-instinct-mi50": {
   "slug": "amd-instinct-mi50",
   "name": "AMD Instinct MI50",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 32,
   "bandwidth_gbs": 1024,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 450.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 117.94,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-rx-6900-xt": {
   "slug": "amd-radeon-rx-6900-xt",
   "name": "AMD Radeon RX 6900 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 512,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 412.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 108.0,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 7
  },
  "amd-radeon-pro-vii": {
   "slug": "amd-radeon-pro-vii",
   "name": "AMD Radeon Pro VII",
   "vendor": "AMD",
   "kind": "workstation",
   "memory_gb": 16,
   "bandwidth_gbs": 1024,
   "spec_source": "Wikipedia (Radeon Pro)",
   "spec_url": "https://en.wikipedia.org/wiki/Radeon_Pro",
   "spec_note": null,
   "eff_bw": 414.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 108.45,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "amd-radeon-rx-6800": {
   "slug": "amd-radeon-rx-6800",
   "name": "AMD Radeon RX 6800",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 512,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 407.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 106.49,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-vii": {
   "slug": "amd-radeon-vii",
   "name": "AMD Radeon VII",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 1024,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 423.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 110.86,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6800-xt": {
   "slug": "amd-radeon-rx-6800-xt",
   "name": "AMD Radeon RX 6800 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 512,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 384.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 100.5,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "amd-radeon-pro-w6800x-duo": {
   "slug": "amd-radeon-pro-w6800x-duo",
   "name": "AMD Radeon Pro W6800X Duo",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 382.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 100.08,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-4070": {
   "slug": "nvidia-rtx-4070",
   "name": "NVIDIA RTX 4070",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 504,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 352.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 92.29,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "nvidia-rtx-3060-ti": {
   "slug": "nvidia-rtx-3060-ti",
   "name": "NVIDIA RTX 3060 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 345.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 90.35,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-pro-w6800x": {
   "slug": "amd-radeon-pro-w6800x",
   "name": "AMD Radeon Pro W6800X",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 330.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 86.47,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-rx-6700-xt": {
   "slug": "amd-radeon-rx-6700-xt",
   "name": "AMD Radeon RX 6700 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 384,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 320.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 83.88,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "amd-radeon-rx-6750-xt": {
   "slug": "amd-radeon-rx-6750-xt",
   "name": "AMD Radeon RX 6750 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 432,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 314.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 82.29,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "amd-radeon-pro-v620": {
   "slug": "amd-radeon-pro-v620",
   "name": "AMD Radeon Pro V620",
   "vendor": "AMD",
   "kind": "workstation",
   "memory_gb": 32,
   "bandwidth_gbs": 512,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 312.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 81.78,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "nvidia-rtx-3070": {
   "slug": "nvidia-rtx-3070",
   "name": "NVIDIA RTX 3070",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 300.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 78.71,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-instinct-mi60": {
   "slug": "amd-instinct-mi60",
   "name": "AMD Instinct MI60",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 32,
   "bandwidth_gbs": 1024,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 349.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 91.46,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "apple-m4-max": {
   "slug": "apple-m4-max",
   "name": "Apple M4 Max",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 546,
   "chip": "M4 Max",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": "assumes 40-core bin",
   "eff_bw": 286.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 75.02,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-tesla-t10": {
   "slug": "nvidia-tesla-t10",
   "name": "Nvidia Tesla T10",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 290.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 76.05,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-9060-xt": {
   "slug": "amd-radeon-rx-9060-xt",
   "name": "AMD Radeon RX 9060 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 322,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 282.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 74.01,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-5700-xt": {
   "slug": "amd-radeon-rx-5700-xt",
   "name": "AMD Radeon RX 5700 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 284.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 74.43,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "intel-arc-b580": {
   "slug": "intel-arc-b580",
   "name": "Intel Arc B580",
   "vendor": "Intel",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 456,
   "spec_source": "Wikipedia (Intel Arc)",
   "spec_url": "https://en.wikipedia.org/wiki/Intel_Arc",
   "spec_note": null,
   "eff_bw": 275.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 72.0,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-pro-v540": {
   "slug": "amd-radeon-pro-v540",
   "name": "AMD Radeon Pro V540",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 266.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 69.64,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-pro-w5700": {
   "slug": "amd-radeon-pro-w5700",
   "name": "AMD Radeon Pro W5700",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 272.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 71.3,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-arc-pro-b60": {
   "slug": "intel-arc-pro-b60",
   "name": "Intel Arc Pro B60",
   "vendor": "Intel",
   "kind": "workstation",
   "memory_gb": 24,
   "bandwidth_gbs": 456,
   "spec_source": "intel.com",
   "spec_url": "https://www.intel.com/content/www/us/en/products/sku/243916/intel-arc-pro-b60-graphics/specifications.html",
   "spec_note": "Intel spec page",
   "eff_bw": 269.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 70.54,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-3070-laptop": {
   "slug": "nvidia-rtx-3070-laptop",
   "name": "NVIDIA RTX 3070 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (384-448)",
   "eff_bw": 243.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 63.64,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-vega-64": {
   "slug": "amd-radeon-rx-vega-64",
   "name": "AMD Radeon RX Vega 64",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 484,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 258.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 67.7,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-p102-100": {
   "slug": "nvidia-p102-100",
   "name": "NVIDIA P102-100",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 5,
   "bandwidth_gbs": 440,
   "spec_source": "techpowerup.com",
   "spec_url": "https://www.techpowerup.com/gpu-specs/p102-100.c3100",
   "spec_note": "TechPowerUp (read via search result; TPU blocks automated fetch): corrected 10 GB -> 5 GB (10 GB card is P102-101); 440.3 GB/s",
   "eff_bw": 254.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 66.48,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-bc-250": {
   "slug": "amd-bc-250",
   "name": "AMD BC-250",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 241.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 63.14,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6650-xt": {
   "slug": "amd-radeon-rx-6650-xt",
   "name": "AMD Radeon RX 6650 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 280,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 246.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 64.53,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "nvidia-rtx-4060-laptop": {
   "slug": "nvidia-rtx-4060-laptop",
   "name": "NVIDIA RTX 4060 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (224-256)",
   "eff_bw": 229.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 60.01,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-7600": {
   "slug": "amd-radeon-rx-7600",
   "name": "AMD Radeon RX 7600",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 232.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 60.88,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "nvidia-gtx-1660-ti-laptop": {
   "slug": "nvidia-gtx-1660-ti-laptop",
   "name": "Nvidia GTX 1660 Ti Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 219.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 57.3,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-instinct-mi25": {
   "slug": "amd-instinct-mi25",
   "name": "AMD Instinct MI25",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 16,
   "bandwidth_gbs": 484,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 244.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 63.94,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6600-xt": {
   "slug": "amd-radeon-rx-6600-xt",
   "name": "AMD Radeon RX 6600 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 206.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 53.92,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-7600-xt": {
   "slug": "amd-radeon-rx-7600-xt",
   "name": "AMD Radeon RX 7600 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 214.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 56.11,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 5
  },
  "intel-arc-a770": {
   "slug": "intel-arc-a770",
   "name": "Intel Arc A770",
   "vendor": "Intel",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 560,
   "spec_source": "Wikipedia (Intel Arc)",
   "spec_url": "https://en.wikipedia.org/wiki/Intel_Arc",
   "spec_note": "16 GB model",
   "eff_bw": 202.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 53.07,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-firepro-s9300-x2": {
   "slug": "amd-firepro-s9300-x2",
   "name": "AMD FirePro S9300 x2",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 212.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 55.64,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-ai-maxplus-395-radeon-8060s": {
   "slug": "amd-ryzen-ai-maxplus-395-radeon-8060s",
   "name": "AMD Ryzen AI Max+ 395 (Radeon 8060S)",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": 128,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of AMD Ryzen processors)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_Ryzen_processors",
   "spec_note": "LPDDR5X-8000, 256-bit, up to 256 GB/s; max 128 GB",
   "eff_bw": 207.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 54.38,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 6
  },
  "amd-radeon-rx-6600": {
   "slug": "amd-radeon-rx-6600",
   "name": "AMD Radeon RX 6600",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 224,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 203.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 53.17,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-vega-56": {
   "slug": "amd-radeon-rx-vega-56",
   "name": "AMD Radeon RX Vega 56",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 410,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 201.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 52.66,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-arc-b570": {
   "slug": "intel-arc-b570",
   "name": "Intel Arc B570",
   "vendor": "Intel",
   "kind": "gpu",
   "memory_gb": 10,
   "bandwidth_gbs": 380,
   "spec_source": "Wikipedia (Intel Arc)",
   "spec_url": "https://en.wikipedia.org/wiki/Intel_Arc",
   "spec_note": null,
   "eff_bw": 193.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 50.49,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-3060-laptop": {
   "slug": "nvidia-rtx-3060-laptop",
   "name": "NVIDIA RTX 3060 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 6,
   "bandwidth_gbs": 336,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (288-336)",
   "eff_bw": 187.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 49.03,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-6800m": {
   "slug": "amd-radeon-rx-6800m",
   "name": "AMD Radeon RX 6800M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 187.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 49.06,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6600m": {
   "slug": "amd-radeon-rx-6600m",
   "name": "AMD Radeon RX 6600M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 184.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 48.21,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-arc-a770m": {
   "slug": "intel-arc-a770m",
   "name": "Intel Arc A770M",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 182.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 47.69,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-arc-pro-b50": {
   "slug": "intel-arc-pro-b50",
   "name": "Intel Arc Pro B50",
   "vendor": "Intel",
   "kind": "workstation",
   "memory_gb": 16,
   "bandwidth_gbs": 224,
   "spec_source": "intel.com",
   "spec_url": "https://download.intel.com/newsroom/2025/client-computing/Intel-Arc-Pro-B50-Data-Sheet.pdf",
   "spec_note": "Intel data sheet",
   "eff_bw": 182.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 47.69,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-p104-100": {
   "slug": "nvidia-p104-100",
   "name": "NVIDIA P104-100",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 4,
   "bandwidth_gbs": 320,
   "spec_source": "gigabyte.com",
   "spec_url": "https://www.gigabyte.com/us/Graphics-Card/GV-NP104D5X-4G/sp",
   "spec_note": "corrected 8 GB -> 4 GB; 256-bit GDDR5X @10 Gbps = 320 GB/s",
   "eff_bw": 185.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 48.64,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-rtx-a2000": {
   "slug": "nvidia-rtx-a2000",
   "name": "NVIDIA RTX A2000",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 6,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 174.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 45.69,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-radeon-rx-7600m-xt": {
   "slug": "amd-radeon-rx-7600m-xt",
   "name": "AMD Radeon RX 7600M XT",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 173.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 45.28,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-pro-v340": {
   "slug": "amd-radeon-pro-v340",
   "name": "AMD Radeon Pro V340",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 181.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 47.54,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-arc-a750": {
   "slug": "intel-arc-a750",
   "name": "Intel Arc A750",
   "vendor": "Intel",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 512,
   "spec_source": "Wikipedia (Intel Arc)",
   "spec_url": "https://en.wikipedia.org/wiki/Intel_Arc",
   "spec_note": null,
   "eff_bw": 168.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 43.96,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-gtx-1070": {
   "slug": "nvidia-gtx-1070",
   "name": "NVIDIA GTX 1070",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 165.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 43.33,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 3
  },
  "amd-radeon-rx-470": {
   "slug": "amd-radeon-rx-470",
   "name": "AMD Radeon RX 470",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 165.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 43.35,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-580": {
   "slug": "amd-radeon-rx-580",
   "name": "AMD Radeon RX 580",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 150.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 39.32,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-rx-570": {
   "slug": "amd-radeon-rx-570",
   "name": "AMD Radeon RX 570",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 224,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 153.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 40.18,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-pro-w5500": {
   "slug": "amd-radeon-pro-w5500",
   "name": "AMD Radeon Pro W5500",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 140.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 36.82,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-rx-480": {
   "slug": "amd-radeon-rx-480",
   "name": "AMD Radeon RX 480",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 142.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 37.23,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "apple-m2-ultra": {
   "slug": "apple-m2-ultra",
   "name": "Apple M2 Ultra",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 800,
   "chip": "M2 Ultra",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 819.2",
   "eff_bw": 645.3,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 Coder 7B",
    "quant": "Q8_0",
    "backend": "Metal",
    "tg": 79.68,
    "size_gb": 8.1,
    "source": "https://github.com/ggml-org/llama.cpp/blob/master/benches/mac-m2-ultra/mac-m2-ultra.md",
    "test": "tg32"
   },
   "n_rows": 8
  },
  "nvidia-gtx-980": {
   "slug": "nvidia-gtx-980",
   "name": "NVIDIA GTX 980",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": 224,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 130.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 34.16,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-firepro-w8100": {
   "slug": "amd-firepro-w8100",
   "name": "AMD FirePro W8100",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 112.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 29.52,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6500-xt": {
   "slug": "amd-radeon-rx-6500-xt",
   "name": "AMD Radeon RX 6500 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": 144,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 106.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 27.81,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "apple-m3": {
   "slug": "apple-m3",
   "name": "Apple M3",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M3",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "theoretical 102.4",
   "eff_bw": 100.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 26.39,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-firepro-s10000": {
   "slug": "amd-firepro-s10000",
   "name": "AMD FirePro S10000",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 96.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 25.32,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-core-ultra-300-series": {
   "slug": "intel-core-ultra-300-series",
   "name": "Intel Core Ultra 300 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 93.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 24.45,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-core-ultra-200-series": {
   "slug": "intel-core-ultra-200-series",
   "name": "Intel Core Ultra 200 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 96.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 25.12,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-ai-300-series": {
   "slug": "amd-ryzen-ai-300-series",
   "name": "AMD Ryzen AI 300 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 91.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 23.93,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-6000-series": {
   "slug": "amd-ryzen-6000-series",
   "name": "AMD Ryzen 6000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 81.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 21.26,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "apple-m2-pro": {
   "slug": "apple-m2-pro",
   "name": "Apple M2 Pro",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M2 Pro",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 204.8",
   "eff_bw": 80.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 20.97,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-gtx-1050-ti": {
   "slug": "nvidia-gtx-1050-ti",
   "name": "NVIDIA GTX 1050 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": 112,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 80.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 20.96,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 4
  },
  "amd-ryzen-7000-series": {
   "slug": "amd-ryzen-7000-series",
   "name": "AMD Ryzen 7000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 76.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 20.09,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-8000-series": {
   "slug": "amd-ryzen-8000-series",
   "name": "AMD Ryzen 8000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 81.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 21.28,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-z1-extreme": {
   "slug": "amd-ryzen-z1-extreme",
   "name": "AMD Ryzen Z1 Extreme",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 71.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 18.77,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-ryzen-ai-400-series": {
   "slug": "amd-ryzen-ai-400-series",
   "name": "AMD Ryzen AI 400 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 72.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 19.03,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-steam-deck": {
   "slug": "amd-steam-deck",
   "name": "AMD Steam Deck",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 70.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 18.4,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-firepro-d700": {
   "slug": "amd-firepro-d700",
   "name": "AMD FirePro D700",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 63.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 16.62,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-radeon-pro-wx-4100": {
   "slug": "amd-radeon-pro-wx-4100",
   "name": "AMD Radeon Pro WX 4100",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 63.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 16.56,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "apple-m2": {
   "slug": "apple-m2",
   "name": "Apple M2",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M2",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M2)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M2",
   "spec_note": "theoretical 102.4",
   "eff_bw": 51.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 13.5,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "apple-m1": {
   "slug": "apple-m1",
   "name": "Apple M1",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 68.25,
   "chip": "M1",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": null,
   "eff_bw": 49.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 12.85,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-5000-series": {
   "slug": "amd-ryzen-5000-series",
   "name": "AMD Ryzen 5000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 42.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 10.98,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-core-1100-series": {
   "slug": "intel-core-1100-series",
   "name": "Intel Core 1100 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 40.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 10.58,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-550": {
   "slug": "amd-radeon-rx-550",
   "name": "AMD Radeon RX 550",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 39.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 10.2,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-ryzen-4000-series": {
   "slug": "amd-ryzen-4000-series",
   "name": "AMD Ryzen 4000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 37.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 9.87,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-tesla-k40": {
   "slug": "nvidia-tesla-k40",
   "name": "Nvidia Tesla K40",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 37.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 9.92,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "mediatek-dimensity-9400": {
   "slug": "mediatek-dimensity-9400",
   "name": "MediaTek Dimensity 9400",
   "vendor": "Other",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 34.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 8.92,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-core-ultra-100-series": {
   "slug": "intel-core-ultra-100-series",
   "name": "Intel Core Ultra 100 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 31.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 8.21,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-ryzen-3000-series": {
   "slug": "amd-ryzen-3000-series",
   "name": "AMD Ryzen 3000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 32.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 8.49,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-core-1000-series": {
   "slug": "intel-core-1000-series",
   "name": "Intel Core 1000 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 16.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 4.25,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "intel-core-8000-series": {
   "slug": "intel-core-8000-series",
   "name": "Intel Core 8000 Series",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 12.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 3.35,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "intel-n150": {
   "slug": "intel-n150",
   "name": "Intel N150",
   "vendor": "Intel",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 11.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 2.93,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "cix-cd8180": {
   "slug": "cix-cd8180",
   "name": "CIX CD8180",
   "vendor": "Other",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 21.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 5.51,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-ryzen-5-3000-series": {
   "slug": "amd-ryzen-5-3000-series",
   "name": "AMD Ryzen 5 3000 Series",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 32.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "Vulkan",
    "tg": 8.47,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/10879",
    "test": "tg128"
   },
   "n_rows": 1
  },
  "amd-instinct-mi300x": {
   "slug": "amd-instinct-mi300x",
   "name": "AMD Instinct MI300X",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 192,
   "bandwidth_gbs": 5300,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 890.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 232.92,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-instinct-mi210": {
   "slug": "amd-instinct-mi210",
   "name": "AMD Instinct MI210",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 64,
   "bandwidth_gbs": 1638,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 497.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 130.18,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-pro-w7900": {
   "slug": "amd-radeon-pro-w7900",
   "name": "AMD Radeon Pro W7900",
   "vendor": "AMD",
   "kind": "workstation",
   "memory_gb": 48,
   "bandwidth_gbs": 864,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 487.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 127.43,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-instinct-mi100": {
   "slug": "amd-instinct-mi100",
   "name": "AMD Instinct MI100",
   "vendor": "AMD",
   "kind": "datacenter",
   "memory_gb": 32,
   "bandwidth_gbs": 1229,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 422.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 110.48,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "amd-radeon-880m": {
   "slug": "amd-radeon-880m",
   "name": "AMD Radeon 880M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 61.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 2 7B",
    "quant": "Q4_0",
    "backend": "ROCm",
    "tg": 16.16,
    "size_gb": 3.82,
    "source": "https://github.com/ggml-org/llama.cpp/discussions/15021",
    "test": "tg128"
   },
   "n_rows": 2
  },
  "nvidia-h100-pcie": {
   "slug": "nvidia-h100-pcie",
   "name": "NVIDIA H100 PCIe",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 80,
   "bandwidth_gbs": 2039,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 648.7,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 72.17,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/69",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-a100-80gb-sxm": {
   "slug": "nvidia-a100-80gb-sxm",
   "name": "NVIDIA A100 80GB SXM",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 80,
   "bandwidth_gbs": 2039,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/data-center/a100/",
   "spec_note": "NVIDIA A100 spec table (80GB SXM 2,039 GB/s)",
   "eff_bw": 596.3,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 66.34,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/61",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-pro-6000-blackwell-max-q": {
   "slug": "nvidia-rtx-pro-6000-blackwell-max-q",
   "name": "NVIDIA RTX PRO 6000 Blackwell Max-Q",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 96,
   "bandwidth_gbs": 1792,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "300 W Max-Q variant; same memory as RTX PRO 6000 Blackwell",
   "eff_bw": 227.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 46.21,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1344",
    "test": "LocalScore avg"
   },
   "n_rows": 6
  },
  "nvidia-l40s": {
   "slug": "nvidia-l40s",
   "name": "NVIDIA L40S",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 48,
   "bandwidth_gbs": 864,
   "spec_source": "nvidia.com",
   "spec_url": "https://www.nvidia.com/en-us/data-center/l40s/",
   "spec_note": "NVIDIA spec table",
   "eff_bw": 225.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 45.91,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/84",
    "test": "LocalScore avg"
   },
   "n_rows": 4
  },
  "nvidia-rtx-3080-ti": {
   "slug": "nvidia-rtx-3080-ti",
   "name": "NVIDIA RTX 3080 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 912,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 445.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 90.49,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/120",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-h100-nvl": {
   "slug": "nvidia-h100-nvl",
   "name": "NVIDIA H100 NVL",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 93,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 184.1,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 20.48,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/78",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-5070-ti-laptop": {
   "slug": "nvidia-rtx-5070-ti-laptop",
   "name": "NVIDIA RTX 5070 Ti Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 12,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 326.2,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 36.29,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/3548",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-3070-ti": {
   "slug": "nvidia-rtx-3070-ti",
   "name": "NVIDIA RTX 3070 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 608,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 409.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 83.17,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/122",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-tesla-v100-16gb": {
   "slug": "nvidia-tesla-v100-16gb",
   "name": "NVIDIA Tesla V100 16GB",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 16,
   "bandwidth_gbs": 900,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 476.1,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 52.97,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/775",
    "test": "LocalScore avg"
   },
   "n_rows": 5
  },
  "nvidia-titan-rtx": {
   "slug": "nvidia-titan-rtx",
   "name": "NVIDIA TITAN RTX",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 24,
   "bandwidth_gbs": 672,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 373.0,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 41.5,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1399",
    "test": "LocalScore avg"
   },
   "n_rows": 4
  },
  "nvidia-rtx-5070-laptop": {
   "slug": "nvidia-rtx-5070-laptop",
   "name": "NVIDIA RTX 5070 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 283.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 57.68,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/4230",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-3500-ada-laptop": {
   "slug": "nvidia-rtx-3500-ada-laptop",
   "name": "NVIDIA RTX 3500 Ada Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 12,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 225.2,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 25.06,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/632",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-3070-ti-laptop": {
   "slug": "nvidia-rtx-3070-ti-laptop",
   "name": "NVIDIA RTX 3070 Ti Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 306.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 62.35,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/741",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-3080-laptop": {
   "slug": "nvidia-rtx-3080-laptop",
   "name": "NVIDIA RTX 3080 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 16,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (384-448; 8 or 16 GB)",
   "eff_bw": 261.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 53.08,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/2029",
    "test": "LocalScore avg"
   },
   "n_rows": 4
  },
  "nvidia-rtx-4070-laptop": {
   "slug": "nvidia-rtx-4070-laptop",
   "name": "NVIDIA RTX 4070 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": 256,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (224-256)",
   "eff_bw": 201.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 40.88,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1537",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-2080-super": {
   "slug": "nvidia-rtx-2080-super",
   "name": "NVIDIA RTX 2080 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 496,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 294.3,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 59.8,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/2011",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-l4": {
   "slug": "nvidia-l4",
   "name": "NVIDIA L4",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 24,
   "bandwidth_gbs": 300,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 212.0,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 23.59,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/63",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-4000-sff-ada": {
   "slug": "nvidia-rtx-4000-sff-ada",
   "name": "NVIDIA RTX 4000 SFF Ada",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 20,
   "bandwidth_gbs": 280,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 221.9,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 24.69,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/157",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-a10": {
   "slug": "nvidia-a10",
   "name": "NVIDIA A10",
   "vendor": "NVIDIA",
   "kind": "datacenter",
   "memory_gb": 24,
   "bandwidth_gbs": 600,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 179.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 36.5,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/363",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-2080": {
   "slug": "nvidia-rtx-2080",
   "name": "NVIDIA RTX 2080",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 287.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 58.43,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1476",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-a5000-laptop": {
   "slug": "nvidia-rtx-a5000-laptop",
   "name": "NVIDIA RTX A5000 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 247.9,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 27.58,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1024",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-4060": {
   "slug": "nvidia-rtx-4060",
   "name": "NVIDIA RTX 4060",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 272,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 188.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 38.3,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/137",
    "test": "LocalScore avg"
   },
   "n_rows": 6
  },
  "nvidia-rtx-2070": {
   "slug": "nvidia-rtx-2070",
   "name": "NVIDIA RTX 2070",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 257.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 52.39,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1727",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-rtx-pro-2000-blackwell-laptop": {
   "slug": "nvidia-rtx-pro-2000-blackwell-laptop",
   "name": "NVIDIA RTX PRO 2000 Blackwell Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 198.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 40.4,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1487",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-2060": {
   "slug": "nvidia-rtx-2060",
   "name": "NVIDIA RTX 2060",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": 336,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 187.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 38.08,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/665",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-2000-ada-laptop": {
   "slug": "nvidia-rtx-2000-ada-laptop",
   "name": "NVIDIA RTX 2000 Ada Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 145.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 29.51,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3256",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-a2000-12gb": {
   "slug": "nvidia-rtx-a2000-12gb",
   "name": "NVIDIA RTX A2000 12GB",
   "vendor": "NVIDIA",
   "kind": "workstation",
   "memory_gb": 12,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "same memory bus/clock as RTX A2000 6GB row",
   "eff_bw": 175.8,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 35.73,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/514",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-3050": {
   "slug": "nvidia-rtx-3050",
   "name": "NVIDIA RTX 3050",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 224,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 169.6,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 34.47,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/2294",
    "test": "LocalScore avg"
   },
   "n_rows": 4
  },
  "apple-m5-max-32c": {
   "slug": "apple-m5-max-32c",
   "name": "Apple M5 Max (32-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 460,
   "chip": "M5 Max",
   "gpu_cores": 32,
   "spec_source": "Wikipedia (Apple M5)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M5",
   "spec_note": "32-core GPU bin; previously fell back to 614 (bug)",
   "eff_bw": 306.9,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 34.15,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/3105",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-1000-ada-laptop": {
   "slug": "nvidia-rtx-1000-ada-laptop",
   "name": "NVIDIA RTX 1000 Ada Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 6,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 119.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 24.18,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3883",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-a2000-8gb-laptop": {
   "slug": "nvidia-rtx-a2000-8gb-laptop",
   "name": "NVIDIA RTX A2000 8GB Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 135.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 27.52,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/731",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-rtx-3050-6gb-laptop": {
   "slug": "nvidia-rtx-3050-6gb-laptop",
   "name": "NVIDIA RTX 3050 6GB Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 6,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 119.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 24.2,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1998",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-rtx-pro-1000-blackwell-laptop": {
   "slug": "nvidia-rtx-pro-1000-blackwell-laptop",
   "name": "NVIDIA RTX PRO 1000 Blackwell Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 97.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 19.85,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/2077",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-quadro-rtx-3000": {
   "slug": "nvidia-quadro-rtx-3000",
   "name": "NVIDIA Quadro RTX 3000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 145.0,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 29.47,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3949",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-gtx-1080": {
   "slug": "nvidia-gtx-1080",
   "name": "NVIDIA GTX 1080",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 320,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 99.5,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 20.23,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/110",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "amd-radeon-rx-6950-xt": {
   "slug": "amd-radeon-rx-6950-xt",
   "name": "AMD Radeon RX 6950 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 576,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 30.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 6.14,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1664",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "apple-m6-12c": {
   "slug": "apple-m6-12c",
   "name": "Apple M6 (12-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 170,
   "chip": "M6",
   "gpu_cores": 12,
   "spec_source": "Wikipedia (Apple M6)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M6",
   "spec_note": "Apple-rated peak per Wikipedia",
   "eff_bw": 124.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 25.24,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/4201",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "amd-radeon-rx-7700s": {
   "slug": "amd-radeon-rx-7700s",
   "name": "AMD Radeon RX 7700S",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 22.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 4.56,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3710",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "apple-m4-8c": {
   "slug": "apple-m4-8c",
   "name": "Apple M4 (8-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 120,
   "chip": "M4",
   "gpu_cores": 8,
   "spec_source": "Wikipedia (Apple M4)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M4",
   "spec_note": null,
   "eff_bw": 75.1,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 15.27,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1394",
    "test": "LocalScore avg"
   },
   "n_rows": 3
  },
  "nvidia-rtx-3050-laptop": {
   "slug": "nvidia-rtx-3050-laptop",
   "name": "NVIDIA RTX 3050 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "apple-m3-8c": {
   "slug": "apple-m3-8c",
   "name": "Apple M3 (8-core GPU)",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 100,
   "chip": "M3",
   "gpu_cores": 8,
   "spec_source": "Wikipedia (Apple M3)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M3",
   "spec_note": "theoretical 102.4",
   "eff_bw": 67.4,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 13.7,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3123",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "apple-a18-pro-5-core-gpu": {
   "slug": "apple-a18-pro-5-core-gpu",
   "name": "Apple A18 Pro (5-core GPU)",
   "vendor": "Apple",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 77.9,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 15.83,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/3128",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "amd-radeon-780m": {
   "slug": "amd-radeon-780m",
   "name": "AMD Radeon 780M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 22.5,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 2.5,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/2943",
    "test": "LocalScore avg"
   },
   "n_rows": 18
  },
  "nvidia-rtx-3050-ti-laptop": {
   "slug": "nvidia-rtx-3050-ti-laptop",
   "name": "NVIDIA RTX 3050 Ti Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "nvidia-gtx-1660-super": {
   "slug": "nvidia-gtx-1660-super",
   "name": "NVIDIA GTX 1660 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": 336,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 47.2,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 9.59,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/1595",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-t1000-8gb": {
   "slug": "nvidia-t1000-8gb",
   "name": "NVIDIA T1000 8GB",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 25.7,
   "calibration": {
    "method": "measured",
    "model": "Llama 3.1 8B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 5.23,
    "size_gb": 4.92,
    "source": "https://www.localscore.ai/accelerator/4232",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-gtx-1650-ti": {
   "slug": "nvidia-gtx-1650-ti",
   "name": "NVIDIA GTX 1650 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "nvidia-quadro-p620": {
   "slug": "nvidia-quadro-p620",
   "name": "NVIDIA Quadro P620",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 2,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "nvidia-a40-48q": {
   "slug": "nvidia-a40-48q",
   "name": "NVIDIA A40-48Q",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 48,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 355.2,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 39.52,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/3109",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-pro-6000-blackwell-server-edition": {
   "slug": "nvidia-rtx-pro-6000-blackwell-server-edition",
   "name": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 95,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 161.2,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 17.94,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1524",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-rtx-pro-2000-blackwell": {
   "slug": "nvidia-rtx-pro-2000-blackwell",
   "name": "NVIDIA RTX PRO 2000 Blackwell",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 208.5,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 23.2,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/3816",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-5000-ada-laptop": {
   "slug": "nvidia-rtx-5000-ada-laptop",
   "name": "NVIDIA RTX 5000 Ada Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 147.3,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 16.39,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1357",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-quadro-p5000": {
   "slug": "nvidia-quadro-p5000",
   "name": "NVIDIA Quadro P5000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 133.6,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 14.86,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1106",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-a4500-laptop": {
   "slug": "nvidia-rtx-a4500-laptop",
   "name": "NVIDIA RTX A4500 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 15,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 71.4,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 7.94,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/1268",
    "test": "LocalScore avg"
   },
   "n_rows": 2
  },
  "nvidia-l40s-8q": {
   "slug": "nvidia-l40s-8q",
   "name": "NVIDIA L40S-8Q",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "nvidia-a2": {
   "slug": "nvidia-a2",
   "name": "NVIDIA A2",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 15,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "eff_bw": 50.0,
   "calibration": {
    "method": "measured",
    "model": "Qwen2.5 14B Instruct",
    "quant": "Q4_K_M",
    "backend": "llamafile",
    "tg": 5.56,
    "size_gb": 8.99,
    "source": "https://www.localscore.ai/accelerator/3738",
    "test": "LocalScore avg"
   },
   "n_rows": 1
  },
  "nvidia-rtx-4090-laptop": {
   "slug": "nvidia-rtx-4090-laptop",
   "name": "NVIDIA RTX 4090 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 16,
   "bandwidth_gbs": 576,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "Wikipedia lists a range by laptop TGP/config; value is the max config (448-576)",
   "eff_bw": 377.9,
   "calibration": {
    "method": "spec",
    "factor": 0.656
   },
   "n_rows": 1
  },
  "nvidia-h200": {
   "slug": "nvidia-h200",
   "name": "NVIDIA H200",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 140,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-quadro-rtx-5000": {
   "slug": "nvidia-quadro-rtx-5000",
   "name": "NVIDIA Quadro RTX 5000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-rtx-a1000-laptop": {
   "slug": "nvidia-rtx-a1000-laptop",
   "name": "NVIDIA RTX A1000 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-rtx-5060": {
   "slug": "nvidia-rtx-5060",
   "name": "NVIDIA RTX 5060",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": 448,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 293.9,
   "calibration": {
    "method": "spec",
    "factor": 0.656
   },
   "n_rows": 1
  },
  "nvidia-rtx-5080-laptop": {
   "slug": "nvidia-rtx-5080-laptop",
   "name": "NVIDIA RTX 5080 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 16,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-rtx-a2000-laptop": {
   "slug": "nvidia-rtx-a2000-laptop",
   "name": "NVIDIA RTX A2000 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-rtx-500-ada-laptop": {
   "slug": "nvidia-rtx-500-ada-laptop",
   "name": "NVIDIA RTX 500 Ada Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-quadro-p6000": {
   "slug": "nvidia-quadro-p6000",
   "name": "NVIDIA Quadro P6000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 22,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-quadro-p4000": {
   "slug": "nvidia-quadro-p4000",
   "name": "NVIDIA Quadro P4000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-rtx-2050": {
   "slug": "nvidia-rtx-2050",
   "name": "NVIDIA RTX 2050",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-p106-090": {
   "slug": "nvidia-p106-090",
   "name": "NVIDIA P106-090",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-gtx-1660-ti": {
   "slug": "nvidia-gtx-1660-ti",
   "name": "NVIDIA GTX 1660 Ti",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 6,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 188.9,
   "calibration": {
    "method": "spec",
    "factor": 0.656
   },
   "n_rows": 1
  },
  "nvidia-gtx-1050": {
   "slug": "nvidia-gtx-1050",
   "name": "NVIDIA GTX 1050",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 2
  },
  "nvidia-gtx-1070-max-q": {
   "slug": "nvidia-gtx-1070-max-q",
   "name": "NVIDIA GTX 1070 Max-Q",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 8,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-gtx-1650-super": {
   "slug": "nvidia-gtx-1650-super",
   "name": "NVIDIA GTX 1650 SUPER",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-gt-1030": {
   "slug": "nvidia-gt-1030",
   "name": "NVIDIA GT 1030",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 2,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-quadro-t2000": {
   "slug": "nvidia-quadro-t2000",
   "name": "NVIDIA Quadro T2000",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-t1200-laptop": {
   "slug": "nvidia-t1200-laptop",
   "name": "NVIDIA T1200 Laptop",
   "vendor": "NVIDIA",
   "kind": "laptop",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-gtx-1650": {
   "slug": "nvidia-gtx-1650",
   "name": "NVIDIA GTX 1650",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": 128,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 84.0,
   "calibration": {
    "method": "spec",
    "factor": 0.656
   },
   "n_rows": 1
  },
  "nvidia-gtx-1650-ti-max-q": {
   "slug": "nvidia-gtx-1650-ti-max-q",
   "name": "NVIDIA GTX 1650 Ti Max-Q",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "amd-radeon-760m": {
   "slug": "amd-radeon-760m",
   "name": "AMD Radeon 760M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "nvidia-mx150": {
   "slug": "nvidia-mx150",
   "name": "NVIDIA MX150",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 2,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "amd-radeon-rx-5500-xt": {
   "slug": "amd-radeon-rx-5500-xt",
   "name": "AMD Radeon RX 5500 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 4,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "amd-radeon-740m": {
   "slug": "amd-radeon-740m",
   "name": "AMD Radeon 740M",
   "vendor": "AMD",
   "kind": "apu",
   "memory_gb": null,
   "bandwidth_gbs": null,
   "spec_source": null,
   "spec_url": null,
   "n_rows": 1
  },
  "apple-m1-max": {
   "slug": "apple-m1-max",
   "name": "Apple M1 Max",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 400,
   "chip": "M1 Max",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 409.6",
   "eff_bw": 332.4,
   "calibration": {
    "method": "spec",
    "factor": 0.831
   },
   "n_rows": 2
  },
  "apple-m1-pro": {
   "slug": "apple-m1-pro",
   "name": "Apple M1 Pro",
   "vendor": "Apple",
   "kind": "mac",
   "memory_gb": null,
   "bandwidth_gbs": 200,
   "chip": "M1 Pro",
   "gpu_cores": null,
   "spec_source": "Wikipedia (Apple M1)",
   "spec_url": "https://en.wikipedia.org/wiki/Apple_M1",
   "spec_note": "theoretical 204.8",
   "eff_bw": 166.2,
   "calibration": {
    "method": "spec",
    "factor": 0.831
   },
   "n_rows": 1
  },
  "nvidia-rtx-4060-ti-16gb": {
   "slug": "nvidia-rtx-4060-ti-16gb",
   "name": "NVIDIA RTX 4060 Ti 16GB",
   "vendor": "NVIDIA",
   "kind": "gpu",
   "memory_gb": 16,
   "bandwidth_gbs": 288,
   "spec_source": "Wikipedia (List of Nvidia graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units",
   "spec_note": "16 GB variant, same 128-bit 18 GT/s bus",
   "eff_bw": 188.9,
   "calibration": {
    "method": "spec",
    "factor": 0.656
   },
   "n_rows": 0
  },
  "amd-radeon-rx-7700-xt": {
   "slug": "amd-radeon-rx-7700-xt",
   "name": "AMD Radeon RX 7700 XT",
   "vendor": "AMD",
   "kind": "gpu",
   "memory_gb": 12,
   "bandwidth_gbs": 432,
   "spec_source": "Wikipedia (List of AMD graphics processing units)",
   "spec_url": "https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units",
   "spec_note": null,
   "eff_bw": 311.9,
   "calibration": {
    "method": "spec",
    "factor": 0.722
   },
   "n_rows": 0
  }
 }
}