Researchers reverse-engineered Apple's M4 Neural Engine, discovering it's a graph execution engine rather than a traditional processor, exposing hidden APIs that bypass CoreML, and revealing that Apple's '38 TOPS' spec is misleading.
A researcher published a reverse engineering analysis of the Apple M4 chip's Neural Engine, revealing its CoreML-based architecture, 6.6 FLOPS/W energy efficiency, and the ability to completely shut down when idle.
A developer with a Mac Mini M4 used Claude to reverse engineer Apple's private Neural Engine APIs, bypassed CoreML, and successfully trained a 110M parameter Microgpt model entirely on the ANE — opening new possibilities for NPU-based AI training.