TL;DR: llama.cpp has re-enabled and extended ESIMD kernels on the FFN path, improving performance for Intel integrated GPUs.
Summary: The latest llama.cpp update (b11718) restores ESIMD kernels on the Feed-Forward Network (FFN) path. This re-enables and extends ESIMD fusion to include Q8_0, specifically benefiting Intel integrated GPUs.
Why it matters: This update can significantly boost inference speed for llama.cpp users leveraging Intel integrated GPUs. Developers should test performance with the new release, especially for quantized models like Q8_0.
Source: github_releases