DigestAI news desk

Cut through the AI noise.

Hardware & Compute1 min read

Piotr Wilkin adds int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4

Piotr Wilkin committed a new int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 in the llama.cpp repository. The change adds a quantized matmul shader that supports multiple quantization formats such as IQ4XS, q80, q41, q50, q51, q3k, q4k, q5k, q6k, and nvfp4, and introduces performance improvements and bug fixes for RDNA4 tuning. The commit is co‑authored by OpenAI Codex and Claude…

1 source primary source

Key points

  • Piotr Wilkin commits int8 coopmat1 matmul for AMD RDNA3/4
  • Adds support for IQ4XS, q80, q41, q50, q51, q3k, q4k, q5k, q6k, nvfp4
  • Fixes segfaults and improves RDNA4 performance

The update targets AMD GPUs, enabling more efficient integer matrix multiplication for large‑language‑model inference. It also adds support for the new BKSTEP parameter, double buffering, and work‑group scheduling optimizations. The commit notes a 1M context size for the Claude Opus 5 model, indicating its relevance for large‑context inference workloads.

Full story from github.com · via Reddit AI communities primary sourceOpen source ↗

vulkan: int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 (... · ggml-org/llama.cpp@70c4e15

github.com · 24 September 2026

Loading the full article…

This text was published by github.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗

Coverage and discussion

1source
Topics · follow one to build your own front page

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Hardware & Compute

All →

Related stories