Tools / LLM GPU calculator / Llama 3.3 70B
Llama 3.3 70B VRAM requirements
GPU memory for the weights and KV cache at FP16, FP8 and 4-bit, and how many GPUs that takes. Figures from the model's config.json.
Short answer
Llama 3.3 70B needs about 160 GB of GPU memory at FP16/BF16, 81 GB at FP8 and 46 GB at 4-bit to serve one user at an 8K-token context. Its 70.6B parameters take 141.2 GB at 16-bit, and that conversation adds 2.7 GB of KV cache. On NVIDIA H100 80 GB GPUs that is 2 GPUs (tensor parallel 2) at FP16 or FP8; for a desktop, at 4-bit it needs 2 RTX 5090s with a layer split, or one 96 GB Mac Studio.
Updated . KV cache at FP16, plus 10% headroom for activations and the runtime.
GPU memory by precision
| Weights | Weight memory | 1 user, 8K context | 32 users, 8K context | 1 user, 128K context |
|---|---|---|---|---|
| FP16/BF16 | 141.2 GB | 160 GB | 252 GB | 205 GB |
| FP8 | 70.6 GB | 81 GB | 174 GB | 126 GB |
| 4-bit (INT4) | 38.8 GB | 46 GB | 139 GB | 91 GB |
Required memory = (weights + KV cache) ÷ 0.9. An FP8 KV cache halves the cache share; set it in the calculator.
How many GPUs for one user at 8K context
| GPU | Memory | FP16/BF16 | FP8 | 4-bit (INT4) |
|---|---|---|---|---|
| Data center · tensor parallel | ||||
| NVIDIA L40S | 48 GB | 4 GPUs · TP4 | 2 GPUs · TP2 | 1 GPU |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | 2 GPUs · TP2 | 1 GPU | 1 GPU |
| NVIDIA H100 SXM | 80 GB | 2 GPUs · TP2 | 2 GPUs · TP2 | 1 GPU |
| NVIDIA H100 NVL | 94 GB | 2 GPUs · TP2 | 1 GPU | 1 GPU |
| NVIDIA H200 | 141 GB | 2 GPUs · TP2 | 1 GPU | 1 GPU |
| NVIDIA B200 | 180 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B300 | 288 GB | 1 GPU | 1 GPU | 1 GPU |
| AMD Instinct MI300X | 192 GB | 1 GPU | 1 GPU | 1 GPU |
| AMD Instinct MI355X | 288 GB | 1 GPU | 1 GPU | 1 GPU |
| Workstation & consumer · layer split | ||||
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 2 cards | 1 card | 1 card |
| NVIDIA GeForce RTX 3090 | 24 GB | 7 cards | 4 cards | 2 cards |
| NVIDIA GeForce RTX 4090 | 24 GB | 7 cards | 4 cards | 2 cards |
| NVIDIA GeForce RTX 5090 | 32 GB | 5 cards | 3 cards | 2 cards |
| Apple Mac Studio M3 Ultra / M5 Ultra | 96 GB (72 usable) | 3 Macs | 2 Macs | 1 Mac |
| Apple Mac Studio M3 Ultra / M5 Ultra | 256 GB (192 usable) | 1 Mac | 1 Mac | 1 Mac |
| Apple Mac Studio M5 Ultra | 512 GB (384 usable) | 1 Mac | 1 Mac | 1 Mac |
Data center GPUs are counted at a tensor-parallel size (TP) that divides the model’s attention heads, within one 8-GPU server. Workstation and consumer cards are counted for a layer split, which works with any number of cards. Mac Studio counts about 75% of unified memory as usable by the GPU. Capacity only: speed depends on memory bandwidth and the serving stack.
Memory is step one. Throughput, latency, power and the network decide the real build.
Size my actual workloadLlama 3.3 70B architecture
- Parameters
- 70.6B
- Layers with a KV cache
- 80
- Attention heads / KV heads
- 64 / 8
- Head dimension
- 128
- Attention
- Grouped-query attention (8 query heads per KV head)
- Maximum context
- 131,072 tokens
- KV cache per token (FP16)
- 2 × 80 layers × 8 KV heads × 128 head dim × 2 bytes = 327,680 bytes (0.33 MB) per token
Source: meta-llama/Llama-3.3-70B-Instruct config.json on Hugging Face. Parameter counts are the Hub’s count of the repo’s safetensors files, or the model card’s figure.
Questions about Llama 3.3 70B
How much VRAM does Llama 3.3 70B need?
About 160 GB at FP16/BF16, 81 GB at FP8 and 46 GB at 4-bit, for one user at an 8K-token context. That is 141.2 GB, 70.6 GB or 38.8 GB of weights, plus 2.7 GB of KV cache, plus 10% headroom for activations and the runtime. Serving 32 users at 8K each takes 252 GB at FP16 or 174 GB at FP8, because every conversation keeps its own KV cache.
How many H100 GPUs does Llama 3.3 70B need?
At FP16/BF16 it takes 2 GPUs (tensor parallel 2) for one user at 8K and 4 GPUs (tensor parallel 4) for 32 users. At FP8, 2 GPUs (tensor parallel 2) for one user at 8K and 4 GPUs (tensor parallel 4) for 32 users. One user at FP8 takes 1 GPU on H200 (141 GB) and 1 GPU on B200 (180 GB).
Can Llama 3.3 70B run on an RTX 4090 or RTX 5090?
At 4-bit it needs about 46 GB for one user at 8K: 2 RTX 4090s or 2 RTX 5090s. At FP8 (81 GB) it takes 4 RTX 4090s or 3 RTX 5090s. Several cards work with a layer split, as in llama.cpp, which divides the model's layers between them.
Can Llama 3.3 70B run on a Mac Studio?
macOS lets the GPU use about 75% of unified memory by default: about 72 GB of a 96 GB, 192 GB of a 256 GB and 384 GB of a 512 GB Mac Studio. At 4-bit Llama 3.3 70B needs 46 GB for one user at 8K, so it fits one 96 GB Mac Studio; at FP8 (81 GB) it fits one 256 GB Mac Studio.
How large is Llama 3.3 70B's KV cache?
At FP16, 2 × 80 layers × 8 KV heads × 128 head dim × 2 bytes = 327,680 bytes (0.33 MB) per token. It uses grouped-query attention: 8 query heads share each KV head, which makes the cache 8× smaller than full multi-head attention. An 8K-token conversation holds 2.7 GB at FP16, and a 128K-token one holds 43 GB.
What is Llama 3.3 70B's maximum context length?
131,072 tokens, according to its Hugging Face configuration. One user at that full length holds 43 GB of KV cache at FP16, for 126 GB in total with FP8 weights, or 102 GB with an FP8 KV cache as well.
Other models
- Llama 3.1 8B
- Llama 3.1 405B
- Llama 4 Scout
- Llama 4 Maverick
- Qwen3 8B
- Qwen3 14B
- Qwen3 32B
- Qwen3-30B-A3B
- Qwen3-235B-A22B
- Qwen2.5 32B
- Qwen2.5 72B
- DeepSeek-V3 / R1
- gpt-oss-20b
- gpt-oss-120b
- Gemma 3 27B
- Mistral Small 3.2 24B
- Mixtral 8x7B
- Phi-4 14B
- GLM-5.3
- LFM2.5-24B-A2B
- Kolibri-1
Not listed? The LLM GPU calculator sizes any Hugging Face model from its ID, link or config.json.
Need to deploy private AI?
We design it, validate it, source it, deploy it and help run it.
Give us your workload, users, budget and timeline. Within 48 hours we'll give you a validated cloud/on-prem architecture, BOM, expected costs and sourcing options.