metal

8 posts tagged “metal”.

1. We Published a 36% Speedup and Withdrew It the Next Day (2026-09-19)
2. Every GPU Out-of-Memory Error Was One sysctl (2026-09-12)
3. 17 GB of Attention Buffers Nobody Read (2026-09-12)
4. Five Kernel Ideas That Measured Neutral or Worse (2026-09-12)
5. We Put the Neural Engine to Work on Prefill. It Added 19%. (2026-09-11)
6. We Wrote Our Own Inference Engine. It Started at 0.9 tok/s. (2026-09-06)
7. We Evaluated a Faster Inference Engine and Rejected It (2026-09-03)
8. Unfusing a Kernel Made It 1.63x Faster (2026-09-01)