The Neural Engine in Apple chipsets like the M4 is only available for Inference, allowing developers to run pre-trained AI models but not to train new ones. Fortunately, one individual has apparently reverse-engineered the SoC by scaling past its software limitations, keeping everything in the RAM, making everything run buttery smooth and incredibly fast.
Even more impressive about this feat is that unlocking the M4’s true potential was possible without tools like CoreML and Metal
On X, @0x0SojalSec has shared code on a GitHub repository on how the true power of the M4 was unlocked. Since Apple doesn’t give any permission level to communicate with these chipsets, the person managed to communicate with the M4 without using any of Apple’s tools like CoreML, Metal, or leveraging the GPU. Instead, this feat was achieved using a custom MIL (Model Intermediate Language) developed from scratch.
Since the hardware is “locked” down, some crafty techniques needed to be employed. For instance, when the process gets stuck and needs to reset to continue training, the custom MIL uses the command “exec()” to “respawn to continue training.” This helps the program refresh its current state so it can keep learning without crashing.
To speed up the process, @0x0SojalSec says that unlocking the M4 was done entirely without writing to the NAND flash, which would be slower. Since everything was written to RAM, it was lightning fast. By breaking down the software restrictions, the M4 in the iPad or Mac can reach AI processing performance of 15.8TFLOPS, which is sufficient for training an AI model without the need to purchase an expensive computer or a ludicrously priced NVIDIA GPU.
Given that this achievement was performed on the M4, imagine the level of performance that can be unlocked on the M5. Unfortunately, we cannot confirm if the same custom MIL can be used on the newer Apple Silicon and if the exec() would perform as expected on the newer platform. Hopefully, we’ll get to see the code in action again. If you want to try it out, click on the link below and share your thoughts in the comments.
News Source: @0x0SojalSec
About the author: Omar Sohail is a reporter and analyst for Wccftech's mobile section, specializing in the technology and business of the mobile industry. His expertise lies in the intricate hardware supply chain, covering developments in semiconductor manufacturing, chip lithography, and camera sensor technology.
Follow Wccftech on Google to get more of our news coverage in your feeds.
