AI Briefingllama.cpp merges dedicated Vulkan Flash Attention prefill kernel for Intel Arc and Xe
·2026.11.13 04:35
Key point
The new kernel leverages native matrix core shapes and SLM architecture, delivering significant speedups on ARL-H, PTL-H, and B70 Pro platforms.
Details
- Merged a dedicated Vulkan Flash Attention prefill kernel for Intel Arc and Xe in llama.cpp master, utilizing customized shaders for native matrix core shapes (m16n16k16/SIMD32) and SLM architecture.
- Achieved massive prefill acceleration on supported platforms, with qwen3 8B Q4_K on ARL-H Windows increasing from 126.58 t/s to 415.31 t/s, and B70 Pro Windows qwen3 8B Q4_K rising from 1064.67 t/s to 2955.83 t/s.
- Enabled for specific architectures like ARL-H and Xe2/Xe3, while remaining disabled for others such as the A770 due to inconsistent test results.
- Backend refinements extend benefits to single-KV-head models like MQA and MLA, with Intel Arc B580 benchmarks showing up to +1530.3% improvement for Ling-3.0-tiny (hkv=1) and +897.3% for Gemma4 E4B Q4_K_M.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
ggml v1.0.0 Released with FP8 Support and Backend OptimizationsHuawei releases MACaQ, a model-agnostic quantizer for local inference