deepseek-v4.1POST

DeepSeek V4.1 Flash on V100 32GB: 17.37 tok/s decode with two cards

Tested on a dual 2650v2, 512GB DDR3 platform: two V100s give 17.37 tok/s decode and 256.29 tok/s prefill, three V100s reach 26.05 and 393.53 tok/s, with the author suspecting an AVX1 bottleneck.

dextune

@dextuneArchived 6 days agoOriginally posted 2026.10.25

dextune @dextuneOriginally posted 2026.10.25

Open original on X

DeepSeek V4.1 Flash 1017 w/ Nvidia V100 32GB CPU : 2650v2 x 2 RAM : DDR3 512GB V100 × 2 Decode 17.37 tps Prefill 256.29 tps V100 × 3 Decode 26.05 tps Prefill 393.53 tps Feels like it could push a bit more TPS. Is it bottlenecked by AVX1? Solid attempt though.

This post quotes another post: “DeepSeek V4.1 Flash IQ2 quant on 3x V100: 26 tok/s decode”

x.com/dextune/status/2114153552558092776Archived 2026.10.25 11:59