GPUs & Macs › AMD Instinct MI50
AMD Instinct MI50: local LLM speed
Llama 2 7B (Q4_0) generates 118 tokens/s on the AMD Instinct MI50 (Vulkan), and processes prompts at 1,127 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 118 | 1,127 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 109 | 1,120 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 106 | 1,129 | Q4_0 | ROCm FA on | tg128 | llama.cpp #15021 details |
| Llama 2 7B | 99.0 | 1,057 | Q4_0 | ROCm FA off | tg128 | llama.cpp #15021 details |
| Qwen2.5 14B Instruct | 12.9 | 48 | Q4_K_M | llamafile | LocalScore avg | LocalScore details |
What fits on the AMD Instinct MI50
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 118 tok/s (Vulkan, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | Q8_0 | 29.0 GB | 15.5 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 35.6 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 47.3 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 101 estimate |
| Qwen3.6-35B-A3B | UD-Q6_K | 29.3 GB | 89.7 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 54.6 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 47.2 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 55.0 estimate |
| Muse-Glimmer-30B | Q8_0 | 29.6 GB | 15.2 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 105 estimate |
| gemma-4-31B-it | Q6_K | 25.2 GB | 17.9 estimate |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 14.2 GB | 103 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 67.1 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 376 estimate |
Similar hardware
AMD Instinct MI60 · AMD Instinct MI100 · AMD Instinct MI25 · AMD Instinct MI210 · AMD Instinct MI300X