NeuroGenesysCan I run Llama 3.1 405B?
Last updated • By Will Schulz
Llama 3.1 405B needs about 261 GB of memory at 4-bit (Q4_K_M) with an 8K-token context, so it needs 4 × 80 GB GPUs, since no single card in our list holds it, or runs on a Mac with 512 GB of unified memory. At 16-bit it needs 841 GB, so it needs more than eight 80 GB GPUs.
How much memory does Llama 3.1 405B need?
| Precision | Weights | Total, 8K context | Total, 32K context | Total, 128K context |
|---|---|---|---|---|
| 16-bit (BF16 / FP16) | 812 GB | 841 GB | 854 GB | 905 GB |
| 8-bit (Q8_0) | 431 GB | 449 GB | 462 GB | 513 GB |
| 4-bit (Q4_K_M) | 248 GB | 261 GB | 273 GB | 324 GB |
Total = weights + FP16 KV cache + 3% of the weights + 1 GB of runtime overhead, as in the VRAM calculator. Longer contexts are only usable up to the model's own context window; check the model card.
Which GPUs can run Llama 3.1 405B?
| GPU memory | 16-bit (BF16 / FP16) | 8-bit (Q8_0) | 4-bit (Q4_K_M) |
|---|---|---|---|
| 8 GB | more than 8 | more than 8 | more than 8 |
| 12 GB | more than 8 | more than 8 | more than 8 |
| 16 GB | more than 8 | more than 8 | more than 8 |
| 24 GBRTX 3090, RTX 4090, L4 | more than 8 | more than 8 | more than 8 |
| 32 GBRTX 5090 | more than 8 | more than 8 | more than 8 |
| 40 GBA100 40GB | more than 8 | more than 8 | 8 cards |
| 48 GBRTX A6000, RTX 6000 Ada, L40S | more than 8 | more than 8 | 8 cards |
| 80 GBA100 80GB, H100 PCIe, H100 SXM | more than 8 | 8 cards | 4 cards |
| 141 GBH200 | 8 cards | 4 cards | 2 cards |
| 180 GBB200 | 8 cards | 4 cards | 2 cards |
| 192 GBAMD MI300X | 8 cards | 4 cards | 2 cards |
Assumes the runtime can allocate 95% of each card and every card pays the fixed overhead. Vendor memory sizes are binary (a 24 GB card holds 25.8 × 10⁹ bytes), and that is accounted for.
Can I run Llama 3.1 405B on a Mac?
| Unified memory | 16-bit (BF16 / FP16) | 8-bit (Q8_0) | 4-bit (Q4_K_M) |
|---|---|---|---|
| 24 GBM4 Pro | no | no | no |
| 32 GBM2 Max | no | no | no |
| 48 GBM3 Max (40-core GPU), M4 Pro, M4 Max (40-core GPU) | no | no | no |
| 64 GBM2 Max, M2 Ultra, M3 Max (40-core GPU), M4 Pro, M4 Max (40-core GPU) | no | no | no |
| 96 GBM2 Max, M3 Ultra | no | no | no |
| 128 GBM2 Ultra, M3 Max (40-core GPU), M4 Max (40-core GPU) | no | no | no |
| 192 GBM2 Ultra | no | no | no |
| 256 GBM3 Ultra | no | no | no |
| 512 GBM3 Ultra | no | no | fits |
macOS lets the GPU use about 75% of unified memory by default; you can raise the limit with sudo sysctl iogpu.wired_limit_mb.
How fast does Llama 3.1 405B run at 4-bit?
| Hardware | Memory bandwidth | Tokens / s | |
|---|---|---|---|
| Apple M3 Ultra512 GB config or larger | 819 GB/s | 2.13 | estimator |
Decoding reads the model weights and the KV cache once per token, so speed is roughly memory bandwidth ÷ bytes per token. Assumes 70% of peak bandwidth on GPUs and 65% on Macs, one GPU or Mac holding the whole model. Only hardware that holds it alone is listed.
Questions
How much VRAM does Llama 3.1 405B need?
About 261 GB at 4-bit (Q4_K_M), 449 GB at 8-bit (Q8_0) and 841 GB at 16-bit (BF16), for one request with an 8K-token context. At 4-bit the weights are 248 GB; the rest is the KV cache and runtime overhead. With 128K tokens of context the 4-bit total rises to 324 GB.
Can I run Llama 3.1 405B on a 24 GB GPU like the RTX 4090?
Not on one card. At 4-bit it needs 261 GB, more than eight 24 GB cards.
Can I run Llama 3.1 405B on a Mac?
Yes: at 4-bit it needs a Mac with at least 512 GB of unified memory, while 16-bit needs more than any Mac we track. By default macOS lets the GPU use about 75% of unified memory, which these figures allow for.
How fast does Llama 3.1 405B run?
At 4-bit, one request generates about 2.13 tokens per second on an Apple M3 Ultra. These are estimates from memory bandwidth (70% of peak on GPUs, 65% on Macs) with 4K tokens in context; real runtimes vary.
Work it out for your setup
- Free toolVRAM calculator: Llama 3.1 405B at Q4_K_MChange the quantization, context, KV-cache precision and batch size.
- Free toolLLM inference speed estimatorTokens per second, time to first token and batching on any GPU or Mac.
- GuideCan I run it? All open-weight modelsMemory needed for every model we track, and what each GPU or Mac can run.
Sources
- Llama 3.1 405B model card and config.json — Hugging Face
