AMD MI355X with FlyDSL kernels beats Blackwell running MiniMax M3.1 in vLLM
AnushElangovan · x · 2026-11-14
Per SemiAnalysis, the Inferact team used FlyDSL kernels to unlock the full potential of AMD's MI355X GPU in vLLM — beating Blackwell on MiniMax M3.1 inference and even surpassing Grace Blackwell at some interactivity levels. Anush Elangovan praised the AMD AI team's work in resharing. A notable "cracked kernel" case showing AMD's inference software stack closing the gap quickly.
More from Infra
- Developer now runs 100% of his agents locally on a Mac Mini, up from 25% earlier this year — josh_bickett · 2026-11-14
- Nvidia's $530B backstop for AI customers: the six assumptions that must hold — iamKierraD · 2026-11-14
- RTX 5070 Ti ComfyUI benchmarks: overclocking and launch flags yield surprising results — AnalElder · 2026-11-14
- Bernstein: Neoclouds earn ~6x more revenue per megawatt than Bitcoin miners — Beth_Kindig · 2026-11-14
- ZONOS3-Mini: MoE TTS model runs realtime on a Pixel 9 and MacBook Air — cocktailpeanut · 2026-11-14
- The Periodic Table of Inference: 12 LLM serving concepts as meme elements — sophiamyang · 2026-11-14