GPUs & Macs › AMD Radeon RX Vega 56
AMD Radeon RX Vega 56: local LLM speed
Llama 2 7B (Q4_0) generates 52.7 tokens/s on the AMD Radeon RX Vega 56 (Vulkan), and processes prompts at 429 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 52.7 | 429 | Q4_0 | Vulkan FA on | tg128 | llama.cpp #10879 details |
| Llama 2 7B | 50.2 | 440 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
What fits on the AMD Radeon RX Vega 56
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 52.7 tok/s (Vulkan, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| gemma-4-12B-it | IQ4_XS | 6.4 GB | 31.6 estimate |
| Qwen3.5-9B | Q5_K_S | 6.4 GB | 31.6 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 44.9 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q4_K_M | 5.8 GB | 34.5 estimate |
| gemma-4-E4B-it | Q5_K_M | 5.5 GB | 36.7 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 47.0 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 168 estimate |
| gemma-4-E2B-it | Q8_0 | 5.0 GB | 40.5 estimate |
| DeepSeek-R1-0528-Qwen3-8B | Q5_K_M | 5.9 GB | 34.4 estimate |
| Llama-3.2-3B-Instruct | Q8_0 | 3.4 GB | 58.8 estimate |
| Phi-4-mini-instruct | Q8_0 | 4.1 GB | 49.3 estimate |
| gemma-3-4b-it | Q8_0 | 4.1 GB | 48.7 estimate |
| Qwen3-14B | UD-IQ3_XXS | 6.0 GB | 33.5 estimate |
| gemma-3-12b-it | Q3_K_M | 6.0 GB | 33.5 estimate |
Similar hardware
AMD Radeon RX 5700 XT · AMD Radeon RX Vega 64 · AMD Radeon RX 6650 XT · AMD Radeon RX 7600 · AMD Radeon RX 6600 XT · AMD Radeon RX 6600 · AMD Radeon RX 580 · AMD Radeon RX 570