GPUs & Macs › AMD Instinct MI100
AMD Instinct MI100: local LLM speed
Llama 2 7B (Q4_0) generates 110 tokens/s on the AMD Instinct MI100 (ROCm), and processes prompts at 2,733 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 110 | 2,733 | Q4_0 | ROCm FA off | tg128 | llama.cpp #15021 details |
| Llama 2 7B | 105 | 2,755 | Q4_0 | ROCm FA on | tg128 | llama.cpp #15021 details |
What fits on the AMD Instinct MI100
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 110 tok/s (ROCm, 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 | 14.5 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 33.3 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 44.3 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 94.2 estimate |
| Qwen3.6-35B-A3B | UD-Q6_K | 29.3 GB | 84.1 estimate |
| gemma-4-26B-A4B-it | Q8_0 | 26.9 GB | 51.1 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 44.2 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 51.5 estimate |
| Muse-Glimmer-30B | Q8_0 | 29.6 GB | 14.3 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 98.7 estimate |
| gemma-4-31B-it | Q6_K | 25.2 GB | 16.8 estimate |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 14.2 GB | 96.3 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 62.9 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 352 estimate |
Similar hardware
AMD Instinct MI50 · AMD Instinct MI60 · AMD Instinct MI25 · AMD Instinct MI210 · AMD Instinct MI300X