llama.cpp Merges Vulkan Flash Attention Kernels for Intel Arc, Boosting Prefill Up to 9-15x
llama.cpp merged dedicated Vulkan Flash Attention prefill kernels for Intel Arc GPUs (PR #29357), leveraging native matrix cores to speed up prefill by up to 9–15x, a major win for local LLM inference on Intel hardware.
2026-11-13 ~ 2026-11-13 · 2 related posts
- llama.cpp merges dedicated Vulkan FA prefill kernel for Intel Arc — up to 15x faster — Polaris_debi5 · 2026-11-13
- llama.cpp merges dedicated Vulkan FA prefill kernel for Intel Arc, up to ~9x faster — Polaris_debi5 · 2026-11-13