Radeon Instinct MI100
AMD
32GB VRAM · 1230 GB/s bandwidth · 184.6 FP16 TFLOPS · 300W TDP
The Radeon Instinct MI100 has 32GB of VRAM with 1230 GB/s memory bandwidth and 184.6 TFLOPS FP16 compute. At Q4 quantization, it can comfortably run Gemma 3 4B (369 tok/s), Qwen 2.5 7B (194 tok/s), Llama 3.1 8B (183 tok/s). Models larger than ~54B parameters won't fit even at Q4. Electricity cost is approximately $32/month at 300W TDP.
What LLMs can you run?
| Model | Params | Q4 Weight | Fit | Decode |
|---|---|---|---|---|
| Gemma 3 4B | 4.0B | 2 GB | comfortable | 369 tok/s |
| Qwen 2.5 7B | 7.6B | 4 GB | comfortable | 194 tok/s |
| Llama 3.1 8B | 8.0B | 4 GB | comfortable | 183 tok/s |
| Mistral Small 24B | 24.0B | 12 GB | comfortable | 61 tok/s |
| Gemma 3 27B | 27.4B | 14 GB | comfortable | 53 tok/s |
| Qwen 2.5 Coder 32B | 32.5B | 16 GB | comfortable | 45 tok/s |
| Llama 3.3 70B | 70.6B | 35 GB | won't fit | |
| Qwen 2.5 72B | 72.7B | 36 GB | won't fit | |
| Llama 3.1 405B | 405B | 202 GB | won't fit | |
| DeepSeek R1 671B | 671B | 336 GB | won't fit |
Similar GPUs
| GPU | VRAM | BW | TFLOPS | TDP |
|---|---|---|---|---|
| Radeon Instinct MI60 | 32GB | 1020 GB/s | 29.5 | 300W |
| TITAN V CEO Edition | 32GB | 868 GB/s | 29.8 | 250W |
| GeForce RTX 5090 | 32GB | 1790 GB/s | 104.8 | 575W |
| GeForce RTX 5090 D | 32GB | 1790 GB/s | 104.8 | 575W |
| Radeon AI PRO R9700 | 32GB | 644 GB/s | 95.7 | 300W |
Compare with another GPU
Select another GPU to compare specs and model performance side by side.