← All resources

Using the Matrix Cores of AMD RDNA 4 architecture GPUs

Auto-discovered

Blog Advanced

Takahiro Harada, Atsushi Yoshimura

Visit resource ↗

Description

Using the Matrix Cores of AMD RDNA 4 architecture GPUs is an AMD GPUOpen technical blog by Takahiro Harada and Atsushi Yoshimura. The page is a single how-to on Wave Matrix Multiply Accumulate (WMMA) for RDNA 4 from HIP kernels.

It compares GEMM FLOPS per clock across RDNA 2 through 4, then explains that RDNA 4 WMMA uses a new VGPR layout with _gfx12 intrinsics and is not backward compatible with RDNA 3. Topics include 16x16 tile GEMM, lane mapping for fragments, packing FP32 pairs into half, dropping RDNA 3 shuffles when chaining matmuls, and a minimal MLP inference example. Sample kernels and an Orochi host program are referenced.

Readers should be advanced GPU programmers comfortable with HIP-style compute and wavefronts. The focus is accelerating matrix math on RDNA 4, not a beginner GPU intro.

The article is free on GPUOpen with inline code and diagrams. Further reading points to the AMD RDNA 4 ISA guide and an earlier RDNA 3 WMMA post.

It helps when porting or writing WMMA kernels so VGPR layouts and intrinsic names match the new generation.