Q: How fast is gpt-oss 20B MoE on the NVIDIA L40S 48GB?

~190–310 tok/s (active-expert estimate) on the NVIDIA L40S 48GB. computed roofline: 864 GB/s × 0.46–0.76 efficiency window / 2.1GB (Q4_K_M) — bands, never points. MoE: file is 14GB (that decides fit) but only ~3B of experts activate per token (~2.1GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math

Question 1

Can the NVIDIA L40S 48GB run gpt-oss 20B MoE?

Accepted Answer

gpt-oss 20B MoE needs about 14.19GB working set (weights 14GB + KV cache 0.19GB at 4k context). The NVIDIA L40S 48GB offers ~48GB usable — verdict: full. 14.2 GB needed of 48 GB usable — headroom for context

Question 2

How fast is gpt-oss 20B MoE on the NVIDIA L40S 48GB?

Accepted Answer

~190–310 tok/s (active-expert estimate)

computed roofline: 864 GB/s × 0.46–0.76 efficiency window / 2.1GB (Q4_K_M) — bands, never points.

MoE: file is 14GB (that decides fit) but only ~3B of experts activate per token (~2.1GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math

Question 3

What should I run gpt-oss 20B MoE on instead?

Accepted Answer

Tesla M40 (24GB, used) runs it fully ($50).

Can the NVIDIA L40S 48GB run gpt-oss 20B MoE?

Run gpt-oss 20B MoE on instead