Observed arrival · 2026-10-31
MacBrains Labs puts Apple’s Neural Engine under the microscope
An independent Apple Silicon research site documents Core ML graph construction, Neural Engine placement, numerical parity, and on-device measurements.
- For
- Apple Silicon and Core ML researchers
- Worth noticing
- The site separates planner-assigned compute placement from hardware tracing, and labels a throughput result whose probe output quality was not validated.
Field notes
The described workflow starts with Swift-authored MILSpec programs, packages and compiles them through Apple's public toolchain, then inspects per-operation assignments and compares outputs against CPU and ANE fixtures. The page reports WikiText-2 perplexity for a 28-layer LUT4 Qwen model run as two 14-layer halves with shared KV and position-correct RoPE. It also documents negative results, including a ring-buffer attention graft that the site says placed nothing on the ANE.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue