A new reverse-engineering effort has uncovered detailed technical specifications of Apple's Neural Engine, the custom silicon powering on-device AI across iPhones, iPads and Macs. The project, dubbed Retrospectively Reverse-Engineering Apple's Neural Engine, provides the most comprehensive public analysis of the coprocessor's instruction set and data flow.
Inside the Neural Engine
Researchers behind the project analyzed firmware images extracted from Apple A-series and M-series chips. They identified a dedicated neural processing unit with a unique instruction set architecture separate from the CPU and GPU. The Neural Engine executes operations such as convolution, pooling and activation functions through 16 parallel compute cores, each capable of processing eight-bit integer data.
The team documented several undocumented instructions that control data movement between the chip's local memory and system RAM. These instructions are critical for optimizing latency and power consumption in real-time AI tasks like speech recognition and image processing.
Why This Matters
The ability to program the Neural Engine directly shifts the balance of power in on-device AI development. Developers can now write low-level kernels that bypass Apple's Core ML framework, potentially achieving higher throughput for custom models. This is especially significant for real-time applications such as augmented reality and autonomous edge computing, where every millisecond counts.
Security researchers, however, now have a clearer target for auditing. The documented instruction set allows them to look for side-channel attacks or timing leaks that could compromise user data processed on the Neural Engine. Apple's tight control over the chip's documentation made such analysis impossible before this reverse-engineering effort.
For the broader AI hardware industry, the project serves as a case study in how proprietary neural processing units can be systematically documented through firmware extraction and static analysis. It may inspire similar efforts for other vendors' accelerators, increasing transparency in the black-box world of dedicated AI chips.
Broader Implications for On-Device AI
The release of this documentation arrives as more machine learning workloads move from the cloud to local devices. Apple's Neural Engine has been a key differentiator for its product line, enabling features like Live Text, FaceID and real-time language translation without sending data to servers. By exposing the engine's inner workings, the project could accelerate the development of third-party applications that exploit this hardware efficiently.
The open-source community has already begun building custom neural network compilers that target the newly documented instruction set. This could lead to a wave of privacy-focused AI applications that run entirely on device with performance competitive with Apple's own implementations. It also raises questions about whether Apple will respond by releasing official programmer's manuals or by hardening the firmware against future analysis.



