# llama.cpp b11623 adds an AMD GPU convolution path

> The release adds a SIMT-based conv2d path for AMD GPUs that lack WMMA support. The project gives no performance figures for the change.

Oossa · 2026-10-14 · https://oossa.com/en/llama-cpp-b11623-adds-an-amd-gpu-convolution-path

The ggml-org team has released llama.cpp b11623, adding a SIMT implicit GEMM path for 2D convolutions on AMD GPUs without WMMA support. The change also skips this path for short reductions; the release notes do not quantify its speed or say which AMD cards benefit. Downloadable builds are listed for macOS, iOS and several Linux platforms.

## The facts

- Release: b11623.
- The AMD conv2d change skips the SIMT path for short reductions.

## Why it matters

For developers running llama.cpp on AMD GPUs, this adds a convolution path when WMMA is unavailable, but the notes leave its performance impact and supported hardware unclear.

## Sources & references

1. [ggml-org/llama.cpp b11623](https://github.com/ggml-org/llama.cpp/releases/tag/b11623) – llama.cpp, 2026-10-14

Last updated: 2026-10-14