New ML Library for Spiking Neural Networks Hits Apple Silicon
A new library called MLX-SNN has been released as the first spiking neural network (SNN) library for Apple's MLX framework. It reportedly runs 2-10x faster and uses less memory than PyTorch alternatives on an M3 Max, achieving up to 97% accuracy on the MNIST dataset.
Apple's MLX framework is a NumPy-like array library specifically designed for Apple Silicon's unified memory architecture. This design eliminates the need for explicit data transfers between the CPU and GPU, a common bottleneck in other frameworks, by allowing both processors to operate on the same memory pool. Spiking Neural Networks (SNNs) represent a "third generation" of neural networks that more closely mimic biological brain function. Instead of continuous activations, they process information using discrete "spikes," making them event-driven and inherently more energy-efficient, which is a significant advantage for low-power and real-time applications on mobile devices. The MLX-SNN library leverages MLX's features like lazy evaluation and composable function transforms to optimize SNN workloads. This native approach is what enables the reported 2.0–2.5x faster training speeds and 3–10x lower GPU memory consumption compared to the PyTorch-based snnTorch library running on identical M3 Max hardware. The library is the first of its kind for the MLX ecosystem and comes equipped with six neuron models (including LIF and Izhikevich), four surrogate gradient functions to enable training, and four spike encoding methods. It also offers a compatible API to make migrating existing SNN code from PyTorch-based frameworks more straightforward. While high-end NVIDIA GPUs remain the industry standard for large-scale model training, Apple Silicon has carved out a niche for local prototyping and inference, particularly for large models that can leverage its unified memory. The development of specialized libraries like MLX-SNN further solidifies the platform's role in the ML research and development workflow. The combination of SNNs' efficiency and a native Apple Silicon framework points toward future applications in robotics, autonomous navigation, and advanced sensor data processing directly on-device. This aligns with the growing need for powerful, efficient AI that can run at the edge, independent of cloud infrastructure.
