Someone reverse-engineered Apple’s Neural Engine. Then trained a 600M parameter model on it.

Apple locked down the ANE for inference only. A weekend project cracked it open for training. The results are real, the limitations are stated up front, and Apple probably isn’t thrilled.

Read the complete technical deep-dive: https://reliable-data-engineering.netlify.app/posts/article_apple_neural_engine_training/

The chip Apple doesn’t want you to touch

Every Mac, iPad, and iPhone sold in the last four years has a chip inside it that almost nobody uses directly.

The Apple Neural Engine. 15.8 TFLOPS of FP16 compute on the M4. That’s serious throughput sitting right there on the die, next to the CPU and GPU. For context, 15.8 TFLOPS is more raw FP16 compute than a 2018-era discrete GPU. It’s a real processor, not a marketing line item.

Apple exposes it through CoreML, but only for inference. Feed a pre-trained model in, get predictions out. Training (the computationally expensive process of actually teaching a model) is explicitly not supported. Apple’s position: use the GPU for that. Or Metal. Or their MLX framework. Just don’t touch the Neural Engine for training. That’s not what it’s for.

This makes a certain amount of sense from Apple’s perspective. The ANE has a narrow, optimized architecture. Keeping the API surface limited means fewer support headaches and fewer developers filing radar bugs about edge cases in hardware they don’t fully understand. But it also means a large fraction of on-die compute goes unused for an entire category of workloads.

A developer who goes by maderix disagreed. Over a series of weekends, they reverse-engineered the private _ANEClient and _ANECompiler APIs, figured out how to compile custom compute graphs at runtime, and got backpropagation running directly on the Neural Engine. The approach required no jailbreaking, no kernel extensions, and no modified system binaries. Everything runs in userspace on a stock macOS installation.

Then they trained a 600-million parameter language model on it.

The repo has 6,000+ stars and three blog posts that walk through the entire journey. The README is unusually candid about what works, what doesn’t, and where the hype outruns the reality.

Want the full breakdown?

Read the complete technical deep-dive: https://reliable-data-engineering.netlify.app/posts/article_apple_neural_engine_training/

Follow Reliable Data Engineering for more technical deep-dives on data engineering, SQL optimization, and AI-assisted development.