ane

An experiment that trains neural networks directly on the Apple Neural Engine using reverse-engineered private APIs. It explores training on hardware normally used only for inference.

Share on XLicense: MIT

Overview

This project is a research experiment that trains neural networks directly on Apple's Neural Engine, which Apple limits to inference through CoreML. It reaches the hardware through reverse-engineered private APIs, _ANEClient and _ANECompiler, without CoreML training, Metal or the GPU. It also documents measured ANE throughput, power and SRAM behavior.

Key features

  • Backpropagation on the Apple Neural Engine
  • Uses reverse-engineered private APIs, no CoreML training or GPU
  • Benchmarks of ANE throughput, power and SRAM behavior
  • Reference for direct ANE access outside CoreML

Best for

Researchers and tinkerers exploring what the Neural Engine can do. The author states it is a research project, not a production framework.

Upstream
maderix/ANE
Fork on GitHub
Guo-astro/ane
Upstream stars
7.3k
Category
Hardware, OS and embedded
Language
Objective-C
License
MIT
Forked
2026-03-02
Sync status
In syncLast synced 2026-09-29