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

Digest AI · Hardware & Compute · published 2026-09-24T13:37:39Z

Canonical: https://digestai.news/story/piotr-wilkin-adds-int8-coopmat1-matmul-implementation-for-amd-rdna3-an

## Summary

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 Opus 5 and includes fixes for segfaults, faster RDNA int‑float conversion, and better cache proximity scheduling.

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.

## 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

## Why it matters

Optimizing int8 matrix multiplication on AMD GPUs can speed up LLM inference, making large‑model deployment more efficient for developers and researchers.

## Sources

1. [vulkan: int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 (... · ggml-org/llama.cpp@70c4e15](https://github.com/ggml-org/llama.cpp/commit/70c4e1582e37e4fd94104eb09301711a0f2675bc) (github.com, 2026-09-24, primary source)

## Cite

Digest AI, "Piotr Wilkin adds int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4", 24 September 2026, https://digestai.news/story/piotr-wilkin-adds-int8-coopmat1-matmul-implementation-for-amd-rdna3-an

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/piotr-wilkin-adds-int8-coopmat1-matmul-implementation-for-amd-rdna3-an.json