llama.cpp merges dedicated Vulkan FA prefill kernel for Intel Arc, up to ~9x faster
Polaris_debi5 · reddit · 2026-11-13
llama.cpp master has merged a dedicated Vulkan Flash Attention prefill kernel for Intel Arc (PR #29357), using custom shaders tuned for Intel's native matrix core shapes (m16n16k16/SIMD32) and SLM architecture. Enabled on ARL-H/Xe2/Xe3, disabled on A770 due to inconsistent results; benefits also extend to single-KV-head models (MQA/MLA). Benchmarks: qwen3 8B prefill jumps from 126.58 to 415.31 t/s on ARL-H and gpt-oss 20B hits 1357 t/s on PTL-H; on B70 Pro qwen3 8B goes 1064→2955 t/s. On an Arc B580, Gemma4 E4B prefill improves +897%, and single-KV-head models like Ling-3.0-tiny go from 76 to 1239 t/s (+1530%).
More from Infra
- Harmonic's Aristotle adds multi-clock domains, latches and clock gating support — satnam6502 · 2026-11-13
- Google, AWS and Microsoft hit by simultaneous outages, sparking BitTorrent jokes — sull · 2026-11-13
- Chris Lattner's PyTorchCon keynote on Mojo and MAX cross-hardware inference is now on YouTube — clattner_llvm · 2026-11-13
- Gavin Baker: China chip curbs may hand AI-powered hardware ecosystem to North Korea, Russia, Iran — GavinSBaker · 2026-11-13
- Did Kimi K3 and DeepSeek-V4-Pro each train on 2,000-3,000 Blackwells? Claim disputed — teortaxesTex · 2026-11-13
- Analyst: Jensen Huang's NVIDIA exit creates a vacuum but not the end of AI buildout — ryanshrout · 2026-11-13