OpenSpike: An OpenRAM SNN Accelerator (2302.01015v1)

Abstract: This paper presents a spiking neural network (SNN) accelerator made using fully open-source EDA tools, process design kit (PDK), and memory macros synthesized using OpenRAM. The chip is taped out in the 130 nm SkyWater process and integrates over 1 million synaptic weights, and offers a reprogrammable architecture. It operates at a clock speed of 40 MHz, a supply of 1.8 V, uses a PicoRV32 core for control, and occupies an area of 33.3 mm^2. The throughput of the accelerator is 48,262 images per second with a wallclock time of 20.72 us, at 56.8 GOPS/W. The spiking neurons use hysteresis to provide an adaptive threshold (i.e., a Schmitt trigger) which can reduce state instability. This results in high performing SNNs across a range of benchmarks that remain competitive with state-of-the-art, full precision SNNs. The design is open sourced and available online: https://github.com/sfmth/OpenSpike

• The paper introduces an open-source SNN accelerator integrating over 1M synaptic weights and processing 1,024 neurons to achieve 48,262 images per second.
• It employs reprogrammable architecture, strategic parallelization, and innovative hysteresis techniques to overcome memory-backed limitations.
• Results include high energy efficiency (56.8 GOPS/W) and strong classification accuracy on datasets like MNIST (99.12%), FashionMNIST (88.12%), and DVSGesture (92.36%).

Overview of OpenSpike: An OpenRAM SNN Accelerator

The paper presents a detailed exploration of OpenSpike, a spiking neural network (SNN) accelerator developed entirely with open-source electronic design automation (EDA) tools and OpenRAM memory macros. The research navigates the complexities of creating an SNN accelerator on the 130 nm SkyWater process, culminating in a chip that integrates over 1 million synaptic weights. Significantly, the architecture offers reprogrammability and operates at a clock speed of 40 MHz, achieving a throughput of 48,262 images per second. This emerges as a compelling alternative to state-of-the-art, full-precision SNNs while promoting the ease of access and reproducibility typical of open-source projects.

Key Architectural Features

OpenSpike employs a reprogrammable architecture with 1,024 hardware neurons processed one layer at a time. The architecture balances the inherent memory-backed limitations of SNNs by using OpenRAM macros to optimize near-memory compute tooling. By introducing hysteresis via a Schmitt trigger, the adaptive thresholding significantly stabilizes neuron states.

The architecture is modular, featuring elements like the Membrane Potential Arithmetic Unit (MAU), Multiply-Accumulate (MAC) units, and a spike processor. Notably, the MAC units are designed to maximize throughput by accumulating inputs four at a time. The overall latency across different stages of network operations is minimized through strategic parallelization and time-multiplexing.

Performance Insights

The accelerator's design showcases strong numerical results, achieving a performance speed of 48,262 images per second with a wallclock time of 20.72 μs. This translates into an energy efficiency of 56.8 GOPS/W, a notable benchmark for low-power applications. Furthermore, the OpenSpike's power consumption is dominated by dynamic power in its combinational components, calculated at 119 mW during peak operations.

In terms of accuracy, OpenSpike demonstrates minimal performance degradation with binarized weights, maintaining high classification accuracy on common image datasets such as MNIST (99.12%), FashionMNIST (88.12%), and DVSGesture (92.36%). These accuracies are impressively close to their full-precision counterparts, affirming the hardware's viability for practical applications.

Implications and Future Directions

The open-source nature of OpenSpike offers significant implications for the field of neuromorphic computing. By lowering the entry barriers, the development of ASICs tailored specifically for neuromorphic workloads could become more prevalent. This not only accelerates algorithmic exploration but also encourages hardware innovation driven by the open-source community.

The use of fully open-source EDA tools and PDKs points towards a promising future where more advanced nodes could be integrated, increasing the accessibility and scalability of neuromorphic initiatives. While current technologies are constrained to more legacy nodes, ongoing industry interest in open-source design flows could foreseeably support processes with higher capabilities.

Given the evolving landscape, OpenSpike serves as a significant step towards realizing the open and detailed exploration of neuromorphic architectures, analogous to the trajectory seen in deep learning methods. Future developments may involve refining power usage, increasing the complexity of supported networks, and extending open-source capabilities to more sophisticated processes, potentially reinforcing the SNN frameworks within varied AI applications.

Through OpenSpike, the persistent dialog between architecture, reproducibility, and open-source implementations continues to shape the modernization and accessibility of neuromorphic research and development. The platform stands as a potential catalyst for the broader dissemination of future-proof, energy-efficient AI technologies.