AMD publishes educational GEMM optimization ladder for Helios MI455X GPUs with HipKittens
salykova_ · x · 2026-09-26
AMD's ROCm blog, together with contributors from Caltech, Stanford, Together AI and the broader community, published an educational GEMM optimization guide for the MI455X (Helios architecture) built on HipKittens.
Key points:
- Helios specs: 432 GB HBM4, 23 TB/s bandwidth, 40 PFLOPs FP4 per GPU; 72 GPUs per rack with 3.6 TB/s scale-up bandwidth, aimed at frontier models and long-context agentic workloads.
- Starting from a baseline kernel, the ladder progressively adds asynchronous memory transfers, Tensor Data Movers (TDM), and workgroup-cluster multicast, showing how each technique shapes performance.
- Inspired by Simon Boehm's classic CUDA GEMM worklog; all code and profiling are public.
A rare high-quality systems tutorial for GPU kernel developers in the AMD ecosystem.
More from Infra
- How Long Until Local ~30B A3B Models Match GLM 5.3 Flash Quality? — Aggravating-Push-207 · 2026-09-26
- Vpipe vs Draw Things on M5 Pro: 24% faster at 1K, finishes 2K where Draw Things crashes — TgoAI · 2026-09-26
- Terafab starts hiring: 1 TW/year chip output and orbital AI compute in its sights — seanmcdonaldxyz · 2026-09-26
- Blog: Scaling LLM Inference from a Single Node to Millions — abhijithneil · 2026-09-26
- Samsung, Oxford and PKU propose TrOPD to distill frontier-model reasoning into on-device small models — jiqizhixin · 2026-09-26
- Pay-as-you-go vs committed LLM API volume: real procurement questions from a scaling team — LeviYagami · 2026-09-26