FlexMan: Design and Implementation of a Hardware Accelerator for Low-Power Neural Networks

FlexMan is a hardware accelerator developed under the CONVOLVE project, designed specifically for low-power Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) to explore energy efficiency optimization solutions for edge AI computing. Its core innovation lies in the flexible computing unit design and dual-mode scheduling mechanism, which can efficiently support both network types on a single architecture, providing a high-energy-efficiency computing solution for edge devices.

With the popularization of AI applications, the demand for neural network computing has grown exponentially, but CPUs/GPUs have insufficient energy efficiency on edge devices, making battery life and heat dissipation bottlenecks. SNNs are naturally low-power due to their event-driven characteristics, but their asynchronous computing nature poses challenges for hardware acceleration.

It adopts a hierarchical storage structure (on-chip SRAM buffer, weight storage unit, neuron state register bank) to reduce energy consumption from external memory access; the parallel multiply-accumulate array supports fixed-point/floating-point operations and is equipped with spike detection logic to handle SNN asynchronous events; dual-mode scheduler: SNN processes spike events sorted by timestamp, while ANN uses an inter-layer pipelining strategy.

Circuit level: Clock gating, power gating technologies, and fine-grained Dynamic Voltage and Frequency Scaling (DVFS) in SNN mode; Storage level: Data compression and sparse coding to reduce the amount of data stored and transmitted; Supports partial inference mode—terminates computation early when confidence is reached to save energy.

Smart homes: Local voice/gesture recognition to protect privacy and reduce bandwidth requirements; Industrial IoT: Equipment status monitoring and predictive maintenance, with local real-time analysis for quick response; Robotics: Running complex perception and decision algorithms on battery-powered platforms, supporting applications like drones and service robots.

Challenges: The SNN programming paradigm differs greatly from traditional frameworks, and there is a lack of user-friendly programming interfaces; Insufficient co-optimization between algorithms and hardware—existing models are mostly designed for GPUs; Process technology: Moore's Law is slowing down, so new devices like memristors and phase-change memory need to be explored.

FlexMan represents an important exploration direction in neural network hardware acceleration, integrating SNNs and ANNs to provide a flexible and efficient solution for edge AI. Its open-source nature provides a foundation for further innovation in academia and industry, and it is worth in-depth research by developers in edge computing and neuromorphic engineering.