Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
Build b10456 fixes a kernel launch bug that throttled quantized inference on Intel GPUs, but CUDA and Metal users get nothing from this release.
Why now: The fix landed in this morning's release, covered in the 2026-08-17 briefing.
llama.cpp shipped build b10456 at 06:29 UTC, fixing a kernel launch bug in the SYCL backend used for Intel GPU inference.
The fix hits one specific group: anyone running local inference on Intel GPUs through SYCL. On an Arc B70, the q4_0 to f32 path jumped from 20.21 to 158.19 GB/s, a 7.8x increase, per the release notes.
The bug lived in the thread and block counts for the quantized copy kernel launch, fixed in PR #27160. Depending on quant size, it under-subscribed or over-subscribed the device, per the release notes. Other quant types on the same hardware stayed flat, which shows the bug was narrow rather than a general SYCL problem.
CUDA and Metal users get nothing from this release. The fix is scoped to the SYCL kernel launch code, so Nvidia and Apple GPU inference stays unchanged.
This fix only touched the q4_0 to f32 path. Check your quant format before expecting a speedup. Other quant types on the same Arc B70 hardware showed no change. Watch future releases for whether the same audit reaches other quant paths on SYCL.
Each link below shares sources, entities, or timing with this story.
Claude Fable uses CUDA / Shared entity: CUDA / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Fable uses CUDA); both cover CUDA; overlapping topics (cuda, kernel, launch).
Shared entities / Same source domain
Both cover CUDA, GitHub, Metal; reported by the same outlet (github.com).
Swift uses Metal / Shared entity: CUDA / Earlier coverage
Linked by a graph relationship (Swift uses Metal); both cover CUDA; earlier CUDA coverage from 2026-06-26.
Shared entities / Same source domain / Earlier coverage / Tension
Both cover CUDA, GitHub; reported by the same outlet (github.com); earlier CUDA coverage from 2026-02-22.
Swift uses Metal / Shared entities / Earlier coverage
Linked by a graph relationship (Swift uses Metal); both cover CUDA, Metal; earlier CUDA coverage from 2026-07-20.
Shared entity: CUDA / Same source domain / Shared topic / Earlier coverage / Downstream implication
Both cover CUDA; reported by the same outlet (github.com); overlapping topics (gpus, llama).
Swift uses Metal / Shared entity: GitHub / Same source domain / Earlier coverage
Linked by a graph relationship (Swift uses Metal); both cover GitHub; reported by the same outlet (github.com).
Swift uses Metal / Shared entity: Metal / Same source domain / Earlier coverage
Linked by a graph relationship (Swift uses Metal); both cover Metal; reported by the same outlet (github.com).