Defending the Apple Neural Engine (ANE) (dennisforbes.ca)

In a recent discussion on Hacker News regarding Apple Research's open-source machine learning framework, MLX, the Apple Neural Engine (ANE) faced disparagement, with some deeming it an overrated feature. The ANE, introduced in 2017 with the iPhone X, is designed specifically for efficient neural network inference, enabling features such as Face ID and various OS functionalities. Over the years, Apple has enhanced the ANE's performance, achieving up to 35 TOPS, which allows it to perform low-power neural operations crucial for practical applications, like text recognition in photos and voice processing through Siri. Critics often overlook the practical applications of ANE, assuming its value is diminished because MLX does not utilize it. The discussion highlights a key point: while the ANE excels in powering specific OS features, it was never intended for heavy-duty neural network training, which Apple addresses with its new tensor cores in recent chips. These new cores are optimized for complex model processing at the expense of higher power consumption. Apple continues to support CoreML for developers who wish to leverage ANE, emphasizing that MLX and CoreML serve different purposes in the AI landscape. This discourse underscores the evolving nature of AI hardware usage and the importance of context in assessing the value of different technologies within the community.