AMD MI355X vLLM performance on MiniMax M3.1 Flash jumps 4x in under 20 days

AnushElangovan · x · 2026-11-18

SemiAnalysis reports that vLLM performance on AMD MI355X running MiniMax M3.1 Flash improved 4x in under 20 days, thanks to new deterministic AMD kernels written in FlyDSL by Inferact that avoid atomic adds for output accumulation. The TCO chart shows tokens per $1 rising 2.55x–4.04x across the interaction range between Oct 23 and Nov 12, with MI355X ownership at $1.5/chip/hr at hyperscaler volume.

Original post →

More from Infra

Infra channel →