A pretty exciting breakthrough has surfaced for Apple Silicon. A developer has figured out how to bypass restrictions on the M4’s Neural Engine (ANE) to enable on-device AI model training, something Apple has kept locked down so far.
Normally, the Neural Engine in M4 chips is mainly used for running (inference) pre-trained models efficiently. But security researcher and developer @0x0SojalSec announced on X that they successfully reverse-engineered the limitations. They’ve managed to unlock up to 15.8 TFLOPS of compute power for full training workloads, including backpropagation on transformer models.
Instead of relying on Apple’s usual Core ML or Metal frameworks, the team built a custom Model Intermediate Language (MIL) to talk directly to the Neural Engine. They keep everything in RAM to avoid slow storage writes, which makes the process surprisingly fast and stable. They even found a workaround using the exec() command to restart the process when training gets stuck, allowing it to checkpoint and keep going.
The developer has shared the project on GitHub, showing that it’s possible to train neural networks directly on the ANE hardware with minimal extra dependencies. Early tests show impressive results, like completing transformer training steps in just milliseconds on M4 chips.
Apple hasn’t officially enabled or supported training on the Neural Engine (they market it as delivering up to 38 TOPS for inference only). This reverse-engineering effort shows there’s a lot more potential hidden in these chips than most people realized.
This is big news for the AI community because it could turn regular Macs and iPads into much more capable local training machines for smaller models, reducing the need to rely on cloud services. It also sparks interesting discussions about Apple’s tightly controlled hardware-software approach.
It’ll be fascinating to see what else the community discovers as more people experiment with this. For now, it’s a promising step toward more powerful on-device AI experimentation in the Apple ecosystem.
Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!
For more daily updates, please visit our News Section.
(Source: @0x0SojalSec on X)
Comments