Can the RTX PRO 6000 Blackwell 96GB Run Llama 3.1 70B?
Runs
Yes. Llama 3.1 70B at Q4_K_M is a 42.5 GB download (Hugging Face file listing for bartowski/Meta-Llama-3.1-70B-Instruct-GGUF), and the RTX PRO 6000 Blackwell 96GB has 96 GB of VRAM, leaving about 53.5 GB for the KV cache and runtime. Published runs put generation at a median of 30.5 tokens per second across 1 cited measurement. Fastest cited run: llm-tracker.
Buy the RTX PRO 6000 Blackwell 96GB
No new listing for this card is tracked on Amazon today; it mostly trades used.
Published runs for Llama 3.1 70B on the RTX PRO 6000 Blackwell 96GB
| Generation | Quant | Runtime | Context | Source | Date |
|---|---|---|---|---|---|
| 30.5 tok/s | Q4_K_M | llama.cpp | — | llm-trackerlabelled "Llama 3.1 70B" | 2025-09-15 |
More about this pair
How much VRAM does Llama 3.1 70B need?
The Q4_K_M file is 42.5 GB. Add roughly 1.5 GB for the runtime and an ~8K-token context, so about 44 GB of VRAM holds it comfortably. Longer contexts grow the KV cache and need more.
What is the fastest Llama 3.1 70B can generate on the RTX PRO 6000 Blackwell 96GB?
Generation reads every weight once per token, so the RTX PRO 6000 Blackwell 96GB's 1792 GB/s of memory bandwidth divided by the 42.5 GB file sets a ceiling near 42 tokens per second. Real runs land below it; SpecPicks rejects any published figure above it.
Check another pair
See the full matrix → · All RTX PRO 6000 Blackwell 96GB benchmarks →
Llama 3.1 70B on other GPUs
- RTX 4060 — Offload
- RTX 3060 Ti — Offload
- RTX 3060 12GB — Offload
- Intel Arc B580 — Offload
- RTX 4070 SUPER — Offload
- RTX 4060 Ti 16GB — Offload
- RTX 5060 Ti 16GB — Offload
- Intel Arc A770 16GB — Offload
- Radeon RX 9070 XT — Offload
- RTX 4070 Ti SUPER — Offload
- RTX 4080 — Offload
- RTX 5070 Ti — Offload
- RTX 5080 — Offload
- RTX 3090 — Offload
- Radeon RX 7900 XTX — Offload
- RTX 4090 — Offload
- RTX 5090 — Offload
- RTX A6000 48GB — Runs, 14.6 tok/s median
Other models on the RTX PRO 6000 Blackwell 96GB
- Llama 3.2 1B — Runs, 244 tok/s median
- Llama 3.2 3B — Runs
- Mistral 7B — Runs
- Qwen2.5 7B — Runs
- Llama 3.1 8B — Runs, 138 tok/s median
- DeepSeek-R1-Distill-Llama 8B — Runs
- Qwen3 8B — Runs
- Gemma 2 9B — Runs
- Gemma 3 12B — Runs
- Qwen2.5 14B — Runs, 81.9 tok/s median
- DeepSeek-R1-Distill-Qwen 14B — Runs
- Qwen3 14B — Runs
- Phi-4 14B — Runs
- gpt-oss-20b — Runs
- Gemma 3 27B — Runs, 61.5 tok/s median
- Gemma 2 27B — Runs
- Qwen3 30B-A3B — Runs
- Qwen3 32B — Runs, 56 tok/s median
- Qwen2.5 32B — Runs
- DeepSeek-R1-Distill-Qwen 32B — Runs, 64.3 tok/s median
- Llama 3.3 70B — Runs, 32 tok/s median
- DeepSeek-R1-Distill-Llama 70B — Runs, 30.2 tok/s median
How the checks work
The fit is the Q4_K_M file size against the card's VRAM, with 1.5 GB kept for the runtime and an ~8K context. A published run counts only if the pair fits, the run is Q4-class, its source does not call it an estimate, it is under the bandwidth ceiling above, and it is not one figure pasted across many cards. The whole matrix is downloadable as a CSV under CC BY 4.0.
As an Amazon Associate, SpecPicks earns from qualifying purchases.
This page is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.