Q: How fast is mixtral 8x7B MoE on the NVIDIA H100 NVL 94GB?

~460–760 tok/s (active-expert estimate) on the NVIDIA H100 NVL 94GB. computed roofline: 3900 GB/s × 0.46–0.76 efficiency window / 3.9GB (Q4_K_M) — bands, never points. MoE: file is 26GB (that decides fit) but only ~1.05B of experts activate per token (~3.9GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math

Question 1

Can the NVIDIA H100 NVL 94GB run mixtral 8x7B MoE?

Accepted Answer

mixtral 8x7B MoE needs about 26.5GB working set (weights 26GB + KV cache 0.5GB at 4k context). The NVIDIA H100 NVL 94GB offers ~94GB usable — verdict: full. 26.5 GB needed of 94 GB usable — headroom for context

Question 2

How fast is mixtral 8x7B MoE on the NVIDIA H100 NVL 94GB?

Accepted Answer

~460–760 tok/s (active-expert estimate)

computed roofline: 3900 GB/s × 0.46–0.76 efficiency window / 3.9GB (Q4_K_M) — bands, never points.

MoE: file is 26GB (that decides fit) but only ~1.05B of experts activate per token (~3.9GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math

Question 3

What should I run mixtral 8x7B MoE on instead?

Accepted Answer

Mac mini M5 Pro (64GB) runs it fully ($1669).

Can the NVIDIA H100 NVL 94GB run mixtral 8x7B MoE?

Run mixtral 8x7B MoE on instead