Best GPU for Llama 3.1 405B: VRAM Requirements
Llama 3.1 405B needs about 895.7 GB of GPU memory in BF16 and 227.4 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
Llama 3.1 405B needs about 895.7 GB of VRAM in FP16, 450.2 GB in FP8 and 227.4 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on 4 × NVIDIA B300 288GB; a 4-bit build fits on one NVIDIA B300 288GB.
Llama 3.1 405B is a 405B-parameter language model from Meta, released in July 2024, with a 131,072-token context window and published under the Llama 3.1 Community License. Its weights alone take 810 GB in BF16 and 202.5 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.
Recommended GPUs for Llama 3.1 405B
Best for inference
2 GPUs × NVIDIA B300 288GB
FP8 or INT8 weights, 450.2 GB at 8,192 tokens.
Best for fine-tuning
8 GPUs × NVIDIA B200 180GB
LoRA on BF16 weights, 1,053 GB.
Budget pick
4 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB
Smallest card class that holds a 4-bit build, 227.4 GB.
Llama 3.1 405B VRAM requirements by precision
One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.
| Precision | Weights | KV cache | Total for inference |
|---|---|---|---|
| BF16 / FP16 | 810 GB | 4.2 GB | 895.7 GB |
| FP8 | 405 GB | 4.2 GB | 450.2 GB |
| INT4 | 202.5 GB | 4.2 GB | 227.4 GB |
Llama 3.1 405B VRAM by task
Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.
| Task | GPU memory |
|---|---|
| Inference at the full 131,072-token context (FP8) | 519.9 GB |
| QLoRA (4-bit base) | 263.3 GB |
| LoRA (BF16 base) | 1,053 GB |
| Full fine-tuning (BF16, AdamW) | 7,128 GB |
GPUs that fit Llama 3.1 405B
Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.
| GPU | Memory | BF16 inference (895.7 GB) | FP8 inference (450.2 GB) | INT4 inference (227.4 GB) | QLoRA (263.3 GB) | LoRA (1,053 GB) |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 32GB | 32 GB | — | — | 8 GPUs | — | — |
| NVIDIA Tesla V100 32GB | 32 GB | — | — | 8 GPUs | — | — |
| NVIDIA A100 PCIe 40GB | 40 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA L40S 48GB | 48 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA Quadro RTX 8000 48GB | 48 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA RTX 6000 Ada 48GB | 48 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA RTX A6000 48GB | 48 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA A100 PCIe 80GB | 80 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA A100 SXM 80GB | 80 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA H100 PCIe 80GB | 80 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA H100 SXM 80GB | 80 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA H200 141GB | 141 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA B200 180GB | 180 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA B300 288GB | 288 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA GB300 NVL | 288 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
What is Llama 3.1 405B?
Llama 3.1 405B is the largest dense model of the Llama 3.1 family. Meta also publishes an FP8 version of the instruction-tuned model, which fits on one 8-GPU server with 80 GB cards; the BF16 weights need more than one such server.
- Developer
- Meta
- Released
- July 2024
- Parameters
- 405B
- Context window
- 131,072 tokens
- License
- Llama 3.1 Community License
Source: Llama 3.1 405B model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does Llama 3.1 405B need?
- For inference with an 8,192-token context and one request, Llama 3.1 405B needs about 895.7 GB in BF16, 450.2 GB in FP8, 227.4 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can Llama 3.1 405B run on a single GPU?
- Yes. In INT4 it needs about 227.4 GB, which fits one NVIDIA B300 288GB (288 GB). Higher precisions need a larger card or several GPUs.
- How much GPU memory does fine-tuning Llama 3.1 405B take?
- QLoRA (4-bit base): about 263.3 GB; LoRA (BF16 base): about 1,053 GB; Full fine-tuning (BF16, AdamW): about 7,128 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for Llama 3.1 405B?
- Best for inference: 2 GPUs × NVIDIA B300 288GB (FP8 or INT8 weights, 450.2 GB at 8,192 tokens). Best for fine-tuning: 8 GPUs × NVIDIA B200 180GB (LoRA on BF16 weights, 1,053 GB). Budget pick: 4 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB (Smallest card class that holds a 4-bit build, 227.4 GB). Check live offers before booking, as availability changes.
- Can Llama 3.1 405B run on an RTX 4090?
- No: even a 4-bit build needs about 227.4 GB, more than eight RTX 4090 cards hold (192 GB). It needs data center GPUs with more memory per card.
- Can Llama 3.1 405B run on an RTX 5090?
- Not on one card: even a 4-bit build needs about 227.4 GB, so it takes 8 RTX 5090 cards (256 GB combined) with the model split across them.
Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.
Explore the live market
Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.