Can I run Llama 3.1 70B on an RTX 5090?
No: Llama 3.1 70B needs about 55.2 GB at Q4_K_M with 32K tokens of context, 23.2 GB more than the 32 GB RTX 5090 holds, so it takes 2 of them (64 GB) in one tensor-parallel group. The smallest setup here that holds Llama 3.1 70B at Q4_K_M with 32K is 2× RTX 5090 (64 GB).
No Q4_K_M with 32K tokens of context on 2 cards
- Q4_K_M, 32K
- 55.2 GB
- RTX 5090
- 32 GB, 1,792 GB/s
- Short by
- 23.2 GB
- Tokens/s
- 33–47 tokens/s
At Q4_K_M with 32K tokens of context on 2 cards it writes about 33–47 tokens/s for one request on 2 RTX 5090 cards.
Llama 3.1 70B on the RTX 5090 as the context fills
One request, FP16 KV cache, 0.5 GB plus 10% overhead; a minus sign is memory missing, and tight is less than 0.5 GB free.
| Context | Q4_K_M | Free | Tokens/s | Q8_0 | Free | Tokens/s |
|---|---|---|---|---|---|---|
| 4K | 45.6 GB | −13.6 GB | — | 78.7 GB | −46.7 GB | — |
| 8K | 47.0 GB | −15.0 GB | — | 80.0 GB | −48.0 GB | — |
| 16K | 49.7 GB | −17.7 GB | — | 82.8 GB | −50.8 GB | — |
| 32K | 55.2 GB | −23.2 GB | — | 88.3 GB | −56.3 GB | — |
| 64K | 66.2 GB | −34.2 GB | — | 99.3 GB | −67.3 GB | — |
| 128K (full) | 88.2 GB | −56.2 GB | — | 121 GB | −89.3 GB | — |
Llama 3.1 70B on more than one RTX 5090
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s | Rent per hour |
|---|---|---|---|---|---|
| 2× (64 GB) | Q5_K_M | 29–42 | 54K | 33–47 | $1.38 |
| 4× (128 GB) | Q8_0 | 40–58 | 128K (full) | 59–89 | $2.76 |
Tensor parallel with the combined bandwidth, as in vLLM; llama.cpp splits layers by default and runs at about one card's speed.
Other options
- The next smaller setting, Q3_K_M, takes 46.8 GB at 32K, 14.8 GB over the RTX 5090; it does not fit with 0.5 GB to spare even with 1K tokens.
- Renting two RTX 5090 GPUs costs about $1.38 an hour (median on getdeploying.com, 2026-09-29): $8.1–12 per million tokens at 33–47 tokens/s.
- Smallest setup for Q4_K_M at 32K: 2× RTX 5090 (64 GB).
Run Llama 3.1 70B on the RTX 5090 with llama-server
llama-server -hf bartowski/Meta-Llama-3.1-70B-Instruct-GGUF:Q4_K_M -c 55296Meta-Llama-3.1-70B-Instruct-Q4_K_M.gguf, 42.5 GB, from bartowski/
Questions
Can I run Llama 3.1 70B on an RTX 5090?
No: Llama 3.1 70B needs about 55.2 GB at Q4_K_M with 32K tokens of context, 23.2 GB more than the 32 GB RTX 5090 holds, so it takes 2 of them (64 GB) in one tensor-parallel group. The smallest setup here that holds Llama 3.1 70B at Q4_K_M with 32K is 2× RTX 5090 (64 GB).
How fast is Llama 3.1 70B on an RTX 5090?
At Q4_K_M with 32K tokens of context on 2 cards it writes about 33–47 tokens/s for one request on 2 RTX 5090 cards.
What does a second RTX 5090 change for Llama 3.1 70B?
Two RTX 5090 cards (64 GB in one tensor-parallel group) hold Llama 3.1 70B at Q5_K_M with 32K, and Q4_K_M up to 54K tokens, at about 33–47 tokens/s.
Try other settings in the VRAM calculator or the speed calculator. See also Llama 3.1 70B VRAM requirements, what LLMs an RTX 5090 can run and every pair, or detect your own GPU. Model data checked .