Developer from Bangladesh hacked the Neural Engine in Apple's M4 chip. Through reverse engineering, he unlocked 15.8 TFLOPS of hidden AI training power. He managed to bypass Apple's software limitations using his own Model Intermediate Language, built to bypass the official CoreML and Metal tools.
Developer from Bangladesh hacked the Neural Engine in Apple's M4 chip. Through reverse engineering, he unlocked 15.8 TFLOPS of hidden AI training power. He managed to bypass Apple's software limitations using his own Model Intermediate Language, built to bypass the official CoreML and Metal tools.
Apple’s M4 processors have a ton of computing power for AI, but historically the company has kept this hardware tightly constrained. By default, the neural engine inside the M4 is completely limited to inference tasks (model inference/execution). This means that developers can only use it to run pre-trained AI models, not to fully train new ones from scratch.
However, a developer has managed to circumvent these strict software restrictions by completely reverse engineering the chip to unlock 15.8 TFLOPS of hidden AI computing power. The breakthrough was made possible by researcher Md Ismail Sojal, who recently published code on GitHub detailing exactly how he managed to unlock the M4’s true potential. What’s particularly impressive is that this result was achieved entirely outside of Apple’s official development ecosystem, Notebookcheck writes .
Since Apple doesn’t provide the necessary access levels to directly interact with the Neural Engine when performing such complex tasks, the developer had to find a way to do without standard tools like CoreML, Metal, or even using a GPU. To do this, he created his own intermediate modeling language from scratch. This custom software successfully filled the gap in the toolkit, allowing him to implement full error backpropagation and transformer training directly on the Apple Neural Engine.
Since the default hardware is severely constrained, the developer also had to implement some very inventive workarounds to maintain system stability. For example, if a process hangs during an intensive training phase, the custom language uses a special execute command to actually restart (respawn) that process. This allows the system to update its current state and continue machine learning from the same place without crashing the entire application.
Speed was also a crucial factor in efficiently handling such a heavy workload. To ensure the smoothest possible learning process, the developer tuned the process so that all data was written exclusively to RAM. By completely avoiding the much slower NAND flash memory, the entire operation remained incredibly fast. For anyone using an M4-based Mac or iPad, this exciting workaround proves that the hardware itself is more than capable of handling the complex tasks of AI training, even if Apple officially prefers to keep these specific capabilities locked down.
