maderix/ane

Training neural networks on Apple Neural Engine via reverse-engineered private APIs

View on GitHub ↗Jump to charts ↓

Summary Information

Updated 8 minutes ago
Added to GitGenius on September 1st, 2026
Created on February 28th, 2026
Open Issues & Pull Requests: 31 (+0)
GitHub issues: Enabled
Number of forks: 961
Total Stargazers: 7,253 (+0)
Total Subscribers: 79 (+0)

Repository Insights (GitGenius)

Most active contributors

Sign in to see contributor activity.

Related repositories by overlapping contributors

No overlapping-contributor repos identified yet.

Charts & Analytics

Fetching additional details & charts...

Issue Activity (beta)

Issue API getrepoissuespagesummary failed: 429 Rate limit exceeded. Please try again later.

Detailed Description

ANE is a research project that enables training neural networks directly on Apple's Neural Engine through reverse-engineered private APIs.

The project addresses the restriction Apple places on the Neural Engine through CoreML, which limits the hardware to inference-only use. By reverse-engineering the private _ANEClient and _ANECompiler APIs, the tool demonstrates that the ANE is capable of training workloads when given direct software access. The approach bypasses CoreML's constraints to show what becomes possible when the hardware's full capabilities are unlocked, documenting real performance characteristics including throughput, power consumption, and SRAM behavior through accompanying benchmarks.

Developers considering this project should understand it is explicitly a proof of concept and research artifact, not a production framework or replacement for established training stacks. It suits researchers exploring direct NPU access and those investigating edge AI optimization, but current utilization remains low at five to nine percent of peak capacity, with many element-wise operations still falling back to CPU execution. The project is not intended as a path to training large models on consumer hardware today, nor as a substitute for GPU training beyond small research-scale models. The maintainer has documented honest results including all limitations in accompanying technical articles rather than overstating implications.

Development activity is deliberately constrained by the maintainer's focus on original research in compiler infrastructure for edge AI optimization. Updates arrive when interesting discoveries emerge rather than on a regular cadence, and the project is explicitly not being grown into a large community effort. Bug fixes and benchmark contributions are welcomed, particularly from those with hardware the maintainer does not own, while feature requests are unlikely to receive attention. The codebase is MIT licensed to encourage forking and independent extension rather than centralizing around a single maintained repository.