Apple Neural Engine

Developer revisits Apple Neural Engine reverse-engineering as M5 folds NPU into GPU

A developer who abandoned reverse-engineering Apple's Neural Engine (ANE) three years ago has resumed the project, this time focused on mapping its full internal architecture rather than making it usable via a Linux driver. The renewed effort was prompted by Apple's M5 chip, which emphasizes LLM performance and integrates ANE cores directly into the GPU, signaling the decline of the standalone neural processing unit design that began with the A11 Bionic in 2017.

eiln.github.io · 2026-09-12

Reverse-engineered fix restores lost DRAM bandwidth on Apple's M3 Neural Engine

A developer profiling Apple's M3 Neural Engine found an RTL performance bug that cuts DRAM weight-streaming speed from a normal 45-60 GB/s down to just 17-19 GB/s whenever a model's total weight size lands on an exact multiple of 1 MiB, affecting 7 of the 15 models in the ANEMLL project. By rerouting around the flawed speculative prefetch path in the kernel's DMA engine, the developer restored full throughput, more than doubling token generation speeds for Llama 3.2 1B and Qwen3-8B.

eiln.github.io · 2026-09-10