RTX 3080 vs RTX PRO 6000 Blackwell
Best sourced output tok/s per model (single-stream generation) — engine and version, quantization, context and a source link on every row. Hover a value for the quant.
RTX 30809 sourced runs · best 277.8 tok/s
RTX PRO 6000 Blackwell29 verified runs · best 1271.3 tok/s
Specs and evidence
| RTX 3080 | RTX PRO 6000 Blackwell | |
|---|---|---|
| VRAM | 10 GB | 96 GB |
| Sourced runs on the board | 9 | 29 |
| Models measured | 9 | 23 |
| Best output tok/s | 277.8 tok/s | 1271.3 tok/s |
| Highest measured peak VRAM | 9.7 GB | 72.7 GB |
| Street price (latest snapshot) | — | — |
Head-to-head
On the 2 models measured on both cards, RTX 3080 leads on 0, RTX PRO 6000 Blackwell on 2. The widest gap is on Muse-Glimmer-30B: RTX 3080 6.0 tok/s vs RTX PRO 6000 Blackwell 152.2 tok/s (RTX PRO 6000 Blackwell 2449% faster).
The measured matrix
| Model | RTX 3080 | RTX PRO 6000 Blackwell |
|---|---|---|
| MobileLLM-125M | — | 1271.3 |
| Qwen2.5-0.5B-Instruct | — | 1060.0 |
| MobileLLM-350M | — | 980.8 |
| SmolLM2-360M-Instruct | — | 968.1 |
| Qwen3-0.6B | — | 956.4 |
| Llama-3.2-1B-Instruct | — | 699.9 |
| gemma-3-1b-it | — | 664.3 |
| Llama-3.2-3B-Instruct | — | 489.2 |
| Ling-3.0-tiny | — | 369.1 |
| Qwen3.6-35B-A3B | — | 330.2 |
| Ornith-1.0-35B | — | 309.7 |
| Qwen3.8-27B | 44.4 | 296.0 |
| LFM2.5-8B-A1B | 277.8 | — |
| DeepSeek-Coder-V2-Lite-Instruct | — | 237.3 |
| Qwen3-Coder-Next | — | 174.3 |
| Llama-3.1-8B-Instruct | — | 158.1 |
| Muse-Glimmer-30B | 6.0 | 152.2 |
| Qwen3.8-27B-MTP | — | 112.1 |
| Qwen3.6-27B | — | 101.4 |
| Qwen3-14B | — | 90.8 |
| Kimi-Linear-48B-A3B-Instruct | — | 90.0 |
| Ornith-1.5-9B | 82.8 | — |
| Qwen3.8-27B-Uncensored | 74.1 | — |
| Bonsai-27B | 72.9 | — |
| gemma-4-12B-it-qat | 71.2 | — |
| Ornith-1.5-35B-A3B | 64.4 | — |
| Qwen3.6-27B-MTP | — | 60.1 |
| gpt-oss-20b | 57.1 | — |
| Qwen2.5-Coder-14B-Instruct | — | 50.3 |
| Qwen3-Coder-30B-A3B-Instruct | — | 26.1 |