Abstract

This paper introduces Wenquxing 22A, a novel, low-power, open-source neuromorphic processor designed to accelerate Spiking Neural Networks (SNNs). It solves the energy inefficiency of traditional CPU-accelerator architectures by integrating the SNN calculation unit directly into the general-purpose RISC-V CPU pipeline using custom instructions (RV-SNN V1.0). Experiments demonstrate that Wenquxing 22A achieves a 5.13 times improvement in energy consumption compared to existing accelerator solutions like ODIN, while maintaining high classification accuracy.

Report

Structured Report: RISC-V Toolchain and Agile Development based Open-source Neuromorphic Processor

Key Highlights

• Novel Processor: Introduction of Wenquxing 22A, a low-power, open-source neuromorphic processor specifically optimized for Spiking Neural Networks (SNN).
• Architectural Integration: The design deviates from standard energy-inefficient CPU-accelerator models by integrating the SNN calculation unit directly into the pipeline of a general-purpose CPU.
• Custom ISA: Utilizes customized RISC-V SNN extension instructions (RV-SNN V1.0) to facilitate efficient neuromorphic computation.
• Energy Efficiency: Achieves significant energy savings, demonstrating a 5.13 times reduction in energy expense compared to the accelerator solution ODIN.
• Open Source: The project follows an Agile Development methodology and releases the source code online on Gitee and GitHub.

Technical Details

• Processor Name: Wenquxing 22A.
• Target Application: Acceleration of Spiking Neural Networks (SNN) for tasks like recognition and classification.
• Instruction Set Architecture: Based on RISC-V, incorporating customized extensions termed RV-SNN V1.0.
• SNN Models Implemented: Streamlined Leaky Integrate-and-Fire (LIF) model and binary stochastic Spike-timing-dependent-plasticity (STDP).
• Performance Data (MNIST): Achieved 91.91% classification accuracy (using 1-bit Wenquxing 22A implementation) compared to 85.00% for the 3-bit ODIN online learning baseline.
• Development Philosophy: Emphasizes Agile Development and provides open-source access to the implementation.

Implications

• RISC-V Expansion into AI: This work validates RISC-V’s strength in domain-specific acceleration by demonstrating how custom instructions can efficiently merge AI logic (SNN calculation) directly into the CPU pipeline, enhancing low-power computing capabilities.
• Advancing Neuromorphic Hardware: Wenquxing 22A provides a critical pathway toward highly energy-efficient neuromorphic computing, which is essential for deploying AI models at the edge or in highly constrained environments where energy consumption is paramount.
• Open-Source Contribution: Releasing the hardware source code accelerates innovation and collaboration within the RISC-V and neuromorphic communities, potentially leading to faster refinement and standardization of SNN extensions for the RISC-V ecosystem.