MiniMax says MI355X is nearing B200 for serving its 428B multimodal model
salykova_ · x · 2026-07-23
MiniMax says it is working with hardware vendors to improve inference performance and keep the stack open. In a quoted note, Ryan Lee says AMD MI355X is now close to Nvidia B200 for serving MiniMax-M3.
The company points to two technical blogs covering full-stack ATOM + ATOMesh co-optimization for its 428B MSA sparse MoE multimodal base model. The optimization details include:
- EAGLE3 speculative decoding to reduce generation latency
- MSA-aligned Page-16 KV cache for 1M long context
- Layer-wise FP8 quantization to avoid precision loss in vision/MoE parts
- Distributed P/D separation for large clusters
The post is mainly about serving efficiency, long-context infrastructure, and hardware/software co-design.
Related event: MiniMax Claims AMD MI355X Serving Performance Nears Nvidia B200(2 posts)→
More from Infra
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27