Question 1

Can NVIDIA GH200 96GB run Llama 3.3 70B Instruct?

Accepted Answer

Yes, NVIDIA GH200 96GB can run Llama 3.3 70B Instruct with a B grade (Runs well). Expected decode speed: 67.4 tok/s.

Question 2

How much VRAM does Llama 3.3 70B Instruct need?

Accepted Answer

Llama 3.3 70B Instruct (70B parameters) requires approximately 64.4 GB of memory with Q4_K_M quantization.

Question 3

What is the best quantization for Llama 3.3 70B Instruct?

Accepted Answer

The recommended quantization for Llama 3.3 70B Instruct is Q4_K_M, which balances quality and memory efficiency.

Workload	Grade	Fit	Decode	TTFT	Context
Agentic Coding	B	Runs well	75.9 tok/s	3711 ms	41K
Chat	C	Runs well	75.9 tok/s	1392 ms	13K
Coding	B	Runs well	67.4 tok/s	2870 ms	24K
RAG	B	Runs well	75.9 tok/s	4639 ms	41K
Reasoning	B	Runs well	75.9 tok/s	3015 ms	24K

Quant	Bits	VRAM	Quality	Fit
Q2_K	2	27.3 GB	Low	D35
Q3_K_S	3	34.3 GB	Low	D37
NVFP4	4

Can NVIDIA GH200 96GB run Llama 3.3 70B Instruct?

Memory breakdown