A100 vs RTX Pro 5000 Blackwell for local AI
The A100 fits more: up to gpt-oss 120B, against Llama 3.3 70B on the RTX Pro 5000 Blackwell. The A100 generates about 16% faster on any model both can run, because it has 1555 GB/s of memory bandwidth to the RTX Pro 5000 Blackwell's 1344 GB/s.
Memory and bandwidth
Memory decides which models fit. Bandwidth decides how fast they generate, because every token reads the model's active weights once.
| A100 | RTX Pro 5000 Blackwell | |
|---|---|---|
| Memory | 40 or 80 GB VRAM | 48 GB VRAM |
| Bandwidth | 1555 GB/s | 1344 GB/s |
| Biggest model that fits | gpt-oss 120B | Llama 3.3 70B |
| Best everyday pick | gpt-oss 120B, ~190 tok/s | Qwen3 32B, ~37 tok/s |
Every model, side by side
Estimated generation speed at the largest memory size each was sold in (80 GB and 48 GB).
| Model | A100 | RTX Pro 5000 Blackwell |
|---|---|---|
| Qwen3 4B About 4 GB | ~340 tok/s Runs great | ~300 tok/s Runs great |
| Qwen3 8B About 7 GB | ~160 tok/s Runs great | ~140 tok/s Runs great |
| Gemma 3 12B About 10 GB | ~110 tok/s Runs great | ~91 tok/s Runs great |
| Qwen3 14B About 11 GB | ~92 tok/s Runs great | ~79 tok/s Runs great |
| gpt-oss 20B About 16 GB | ~270 tok/s Runs great | ~240 tok/s Runs great |
| Gemma 3 27B About 19 GB | ~50 tok/s Runs great | ~43 tok/s Runs great |
| Qwen3 30B-A3B About 21 GB | ~270 tok/s Runs great | ~240 tok/s Runs great |
| Qwen3 32B About 22 GB | ~43 tok/s Runs great | ~37 tok/s Runs great |
| Llama 3.3 70B About 45 GB | ~20 tok/s Usable | ~17 tok/s Usable |
| gpt-oss 120B About 67 GB | ~190 tok/s Runs great | Needs 67 GB |
| Qwen3 235B-A22B About 144 GB | Needs 144 GB | Needs 144 GB |
| DeepSeek R1 671B About 406 GB | Needs 406 GB | Needs 406 GB |
The A100: it fits bigger models and runs faster. Price matters as much as either number: check used prices on the builds page, where every parts list has a measured or estimated speed.
Both columns use the same formula: memory bandwidth divided by the weights read per token, calibrated on our own rack, with Ollama's default 4-bit sizes. Real results are usually within about a third. Measured either one? Send us the number.