HipKittens support comes to AMD Instinct MI455X GPUs
AMD says its code walkthrough takes developers from a baseline GPU kernel through optimizations including asynchronous memory transfers and Tensor Data Movers, showing how each technique affects performance.
TLDR
AMD says it worked with contributors from Caltech, Stanford, @togethercompute and the broader developer community to bring HipKittens support to Instinct MI455X GPUs. The company describes the resulting matrix-multiplication (GEMM) kernels for the Helios GPU architecture as “peak-performance.” Its walkthrough starts with a baseline kernel and progresses through asynchronous memory transfers, Tensor Data Movers (TDM) and workgroup-cluster multicast, with code illustrating how each technique affects performance.
Combined views
—
3 Sources, first seen ago
— likes— comments— saves— reposts
