LLMScalelive Hugging Face data
ENTR

Llama 3.3 70B Instruct: how much VRAM does it need?

Llama 3.3 70B Instruct has 70.55B parameters in 80 layers. It uses grouped-query attention: 64 query heads share 8 key-value heads, so its KV cache is 8× smaller than full multi-head attention would need. It takes up to 131,072 tokens of context; one user at that length needs 40.00 GiB of KV cache at FP16 on top of the weights. Its repository is gated on Hugging Face, so LLMScale ships its architecture in a bundled database.

BF16 · 8k context · one user
137.6 GiB

131.4 GiB of weights, 2.50 GiB of KV cache and 3.73 GiB of vLLM's overhead. At INT4 the same workload needs 44.01 GiB.

  • 80 layers
  • hidden 8192
  • 64/8 heads
  • head_dim 128
  • 128k max context

Llama 3.3 70B Instruct memory by precision and context

Precision8k context32k context128k context
BF16137.6 GiBKV cache 2.50 GiB145.1 GiBKV cache 10.00 GiB175.1 GiBKV cache 40.00 GiB
FP873.89 GiBKV cache 2.50 GiB81.39 GiBKV cache 10.00 GiB111.4 GiBKV cache 40.00 GiB
INT444.01 GiBKV cache 2.50 GiB51.51 GiBKV cache 10.00 GiB81.51 GiBKV cache 40.00 GiB

One user, FP16 KV cache and vLLM's overhead on one GPU at its defaults for a card under 70 GB. Quantizing shrinks the weights, not the KV cache, so the saving narrows as the context grows.

Which single GPUs hold it

At 8k context for one user, the smallest cards first. A discrete card is budgeted at vLLM's default 92% of its memory, a unified-memory device at the share its GPU can address, and each carries vLLM's overhead at its own defaults.

How many users fit

GPUBF16FP8INT4
RTX 4090does not fitdoes not fitdoes not fit
RTX 5090does not fitdoes not fitdoes not fit
RTX 6000 Adadoes not fitdoes not fitdoes not fit
H100 SXM 80GBdoes not fitdoes not fit12
RTX PRO 6000 Blackwelldoes not fit618
AMD Ryzen AI Max+ 395 (128GB)does not fit618
H200 141GBdoes not fit2234
B200 180GB103547

Users at 8k context each, on one card. ✗: the weights leave no room for even one.

Fine-tuning Llama 3.3 70B Instruct

MethodMemorySmallest GPU
Full fine-tune1054.8 GiBseveral GPUs, sharded
LoRA (rank 16, attention)135.8 GiBApple M2 Ultra
QLoRA (rank 16, attention)38.26 GiBNVIDIA L40S

Sequence length 2,048, batch size 1, gradient checkpointing on, Adam. Full fine-tuning keeps 16 bytes per parameter before activations; QLoRA stores the frozen base in 4-bit NF4.

Where the numbers come from

Weights are parameters × bytes per parameter. The KV cache is 2 × layers × key-value heads × head dimension × bytes × context × users. For Llama 3.3 70B Instruct that is 320.0 KiB per token at FP16. These are capacity-planning figures: the serving engine — vLLM, TGI, llama.cpp — and its paged-attention efficiency still move the real number.

More on KV cache size and on GQA, MQA and MLA.

Work with Llama 3.3 70B Instruct

Other Llama models