Apple M4 Neural Engine Unlocked for AI Training via Reverse Engineering

A security researcher has reverse-engineered Apple's private APIs to unlock the M4 Neural Engine for AI training. The work bypasses Apple's restrictions, allowing full backpropagation and transformer training directly on the Neural Engine.

Researcher bypasses Apple's inference-only restriction

The M4 chip powers recent Macs and iPads. Apple originally limited the Neural Engine to inference tasks only. The researcher, known as @0x0SojalSec, developed a custom Model Intermediate Language (MIL) to communicate with the M4 hardware.

The unlocked M4 achieves 15.8 TFLOPS of AI processing performance. This was accomplished without using CoreML, Metal, or the GPU. The process uses exec() to respawn training when it stalls, avoiding NAND writes by relying on RAM.

It remains unclear whether the same custom MIL will work on M5 or newer Apple Silicon. The researcher has not confirmed compatibility with future chips. The work is available on GitHub for those interested in replicating the results.