LLM GPU requirements
Llama 3.1 405B GPU requirements
How much GPU memory Llama 3.1 405B needs in BF16, FP8 and 4-bit, how its KV cache grows with context, what fine-tuning takes, and which servers run it, with monthly prices.
- Parameters
- 405B
- Architecture
- Dense 126 layers
- KV heads
- 8 × 128 dim
- Context
- 128k tokens
Cheapest servers that run it
FP8 weights · 8k context · 8 concurrent requests · about 480 GB
-
8× NVIDIA RTX PRO 6000 Blackwell In stock
8 × 96 GB GDDR7 · 8-GPU server · PCIe, no NVLink
$10,490/mo
-
8× NVIDIA HGX H100 In stock
8 × 80 GB HBM3 · 8-GPU server
$16,900/mo
-
8× AMD Instinct MI355X Limited
8 × 288 GB HBM3E · 8-GPU server
$19,900/mo
Delivery date confirmed before you pay. Missed it? Full refund.
Inference memory
How much VRAM Llama 3.1 405B needs
Weights take parameters × bytes per parameter. Serving adds the KV cache and about 10% runtime overhead; the totals below assume an 8k context with 8 concurrent requests and a BF16 KV cache.
| Precision | Weights | Total to serve | Cheapest server that fits |
|---|---|---|---|
| BF16 2 bytes per parameter | 810 GB | 926 GB | 8× AMD Instinct MI355X $19,900/mo |
| FP8 1 byte per parameter | 405 GB | 480 GB | 8× NVIDIA RTX PRO 6000 Blackwell $10,490/mo |
| 4-bit 0.5 bytes per parameter | 203 GB | 258 GB | 8× NVIDIA RTX PRO 6000 Blackwell $10,490/mo |
Estimates with about 8% of GPU memory kept free. Quantized weights also keep some layers in higher precision, and frameworks differ in overhead; check the free memory your engine reports at startup. A model split across GPUs with tensor parallelism runs best on NVLink servers (HGX); the 8× RTX PRO 6000 server links its GPUs over PCIe.
KV cache
Llama 3.1 405B KV cache by context length
Each token in flight stores keys and values for every layer: 2 × 126 layers × 8 KV heads × 128 dimensions × 2 bytes = 504 KB per token in BF16. An FP8 KV cache halves these figures.
| Context per request | 1 request | 8 requests | 32 requests |
|---|---|---|---|
| 4k tokens | 2.1 GB | 17 GB | 68 GB |
| 32k tokens | 17 GB | 135 GB | 541 GB |
| 128k tokens | 68 GB | 541 GB | 2,165 GB |
Fine-tuning
Memory to fine-tune Llama 3.1 405B
Rough totals before tuning batch size and sequence length: QLoRA keeps 4-bit base weights, LoRA keeps BF16 weights frozen, and a full fine-tune with AdamW needs about 16 bytes per parameter plus activations, sharded across GPUs.
| Method | Approximate memory | Cheapest server that fits |
|---|---|---|
| QLoRA (4-bit base) | 306 GB | 8× NVIDIA RTX PRO 6000 Blackwell $10,490/mo |
| LoRA (BF16 base) | 1,004 GB | 8× AMD Instinct MI355X $19,900/mo |
| Full fine-tune (AdamW, mixed precision) | 7,776 GB | Multi-node: GPU clusters |
Full fine-tuning assumes sharded training (FSDP or ZeRO-3) with activation checkpointing. See the fine-tuning guide for methods and settings.
About Llama 3.1 405B
Llama 3.1 405B is Meta’s largest dense open model: 126 layers, 8 KV heads and a 128k-token context window. At about 405 GB in FP8 it needs a full 8-GPU server, and its 126 layers make the KV cache per token the largest of the models on this site.
Every figure on this page comes from the model’s published architecture and the same formulas as our VRAM calculator, where you can change the context length, concurrency and KV cache precision.
Model facts
- Developer
- Meta
- License
- Llama 3.1 Community License
- Type
- Dense
- Parameters
- 405B
- Layers
- 126
- Attention
- 128 query heads, 8 KV headsgrouped-query attention, head dimension 128
- Context window
- 128k tokens
FAQ
Llama 3.1 405B GPU FAQ
The short version. Anything else, our engineers reply 24/7.
Ask an engineerHow much VRAM does Llama 3.1 405B need?
About 810 GB for the weights in BF16, 405 GB in FP8 and 203 GB in 4-bit, plus the KV cache: about 0.5 GB per 1,000 tokens in flight with a BF16 cache. Serving it in FP8 with an 8k context and 8 concurrent requests takes about 480 GB in total.
Can Llama 3.1 405B run on a single GPU?
Not at an 8k context with 8 concurrent requests: in FP8 it needs about 480 GB, more than any single GPU holds. An 8-GPU server such as the 8× NVIDIA RTX PRO 6000 Blackwell runs it with tensor parallelism.
What is the cheapest FullGPU server for Llama 3.1 405B?
For FP8 serving with an 8k context and 8 concurrent requests, the 8× NVIDIA RTX PRO 6000 Blackwell at $10,490 per month is the lowest-priced configuration that fits. For other contexts and concurrency, size it in the VRAM calculator or ask an engineer.
How much memory do I need to fine-tune Llama 3.1 405B?
Roughly 306 GB with QLoRA, 1,004 GB with LoRA on BF16 weights, and 7,776 GB for a full fine-tune with AdamW in mixed precision, before tuning batch size and sequence length. Full fine-tunes of large models are sharded across GPUs with FSDP or ZeRO-3.
Talk to an engineer
Run Llama 3.1 405B on your own GPUs.
Tell us what you need. A GPU engineer replies with availability, a confirmed delivery date and a payment link.
- Delivery date confirmed before you pay
- Cancel any month
- Card or crypto, same price