apple-silicon

23 posts tagged “apple-silicon”.

1. fp16 Residual Streams and a Two-Million Activation (2026-09-25)
2. The Win Was Four Slots, Not the Prefix Cache (2026-09-25)
3. We Published a 36% Speedup and Withdrew It the Next Day (2026-09-19)
4. Our Two Machines Ran Different Models Under the Same Name (2026-09-18)
5. Speculative Decoding Won Alone and Lost Batched (2026-09-18)
6. Your Headline tok/s Is a Short-Prompt Number (2026-09-18)
7. Run Local AI on a Mac: A Realistic Guide (What Works, What Doesn't) (2026-09-17)
8. We indexed 45 local-inference benchmark results — the 4.7x gap nobody looks for (2026-09-14)
9. Every GPU Out-of-Memory Error Was One sysctl (2026-09-12)
10. 17 GB of Attention Buffers Nobody Read (2026-09-12)
11. Can SSD Offload Fit a Model Twice Your RAM? The Arithmetic. (2026-09-12)
12. Five Kernel Ideas That Measured Neutral or Worse (2026-09-12)
13. We Put the Neural Engine to Work on Prefill. It Added 19%. (2026-09-11)
14. We Wrote Our Own Inference Engine. It Started at 0.9 tok/s. (2026-09-06)
15. We Panicked the Same Machine Twice in One Day (2026-09-06)
16. Our Kernel Benchmark Was Measuring the System Cache (2026-09-05)
17. We Evaluated a Faster Inference Engine and Rejected It (2026-09-03)
18. 2x Decode From Block-Diffusion Drafting (2026-09-03)
19. Benchmark the Hardware Before You Benchmark the Software (2026-09-02)
20. Unfusing a Kernel Made It 1.63x Faster (2026-09-01)
21. A 192 GB Model on a 256 GB Box (2026-08-27)
22. One Environment Variable Cost Us 60x on Cache Hits (2026-08-24)
23. One Abandoned Request Wedged the Whole Server (2026-08-23)