Can the NVIDIA RTX 5090 run Llama 3.1 70B?
- FP16
- Q8_0
- Q6_K
- Q5_K_M
- Q4_K_M
- Q3_K_M
- Q2_K
Only just. Llama 3.1 70B fits on the NVIDIA RTX 5090 at Q2_K, which costs noticeable quality. Expect about 41 tokens/second. A card with more memory is a better experience.
Run it with Ollama
Ollama's default llama3.1:70b download is Q4_K_M, which needs about 46.1 GB, more than the 32 GB available. Download the Q2_K GGUF instead (LM Studio lists quants per model), or try a smaller context length.
New to this? Step-by-step setup for the RTX 5090
Run it in the cloud
Want full quality? Run Llama 3.1 70B on a bigger rented GPU.
Rent a cloud GPU by the hour. No upfront cost, stop any time. Not sure whether to rent or buy? Compare the cost.
Memory needed by quantization
At 8,192 tokens of context. The black line marks the 32 GB the NVIDIA RTX 5090 can use.
| Quant | Quality | Memory | Fits? | Speed | |
|---|---|---|---|---|---|
| FP16 | Full quality | 144.5 GB | Too big | — | |
| Q8_0 | Near-lossless | 78.3 GB | Too big | — | |
| Q6_K | Excellent | 61.2 GB | Too big | — | |
| Q5_K_M | Very good | 53.5 GB | Too big | — | |
| Q4_K_M | Good, most popular | 46.1 GB | Too big | — | |
| Q3_K_M | Noticeably degraded | 37.8 GB | Too big | — | |
| Q2_K | Heavily degraded | 29.8 GB | Fits | ~41 tok/s |
GPUs that run Llama 3.1 70B comfortably
Frequently asked questions
Can NVIDIA RTX 5090 run Llama 3.1 70B?
Only just. Llama 3.1 70B fits on the NVIDIA RTX 5090 at Q2_K, which costs noticeable quality. Expect about 41 tokens/second. A card with more memory is a better experience.
How much VRAM does Llama 3.1 70B need?
About 46.1 GB at Q4_K_M with 8,192 tokens of context, and about 144.5 GB at full FP16 precision.
How fast is Llama 3.1 70B on the NVIDIA RTX 5090?
Roughly 41 tokens per second at Q2_K, estimated from the NVIDIA RTX 5090's 1792 GB/s memory bandwidth.