Abstract
SpinAPS is a novel hardware accelerator designed for probabilistic Spiking Neural Networks (SNNs) using spintronic memory (STT-RAM) and digital CMOS logic. By employing a Generalized Linear Model (GLM) for neurons and a "first-to-spike" (FtS) decoding strategy, SpinAPS achieves significant improvements in energy and area efficiency. Experimental results on MNIST and HAR datasets demonstrate that SpinAPS offers a 4× performance boost over SRAM-based designs and 20× better energy efficiency than high-end GPUs, while maintaining competitive accuracy.
Introduction
Modern artificial intelligence is largely driven by Deep Neural Networks (DNNs), yet their deployment on traditional von Neumann architectures is hindered by the "memory wall"—the energy-intensive process of moving data between separate memory and processing units. Spiking Neural Networks (SNNs) offer a bio-inspired solution by using sparse, discrete spikes for communication, which theoretically reduces power consumption. However, implementing SNNs on hardware remains challenging due to training difficulties and the noise inherent in analog memristive synapses.
SpinAPS (Spintronic Accelerator for Probabilistic SNNs) addresses these challenges by utilizing binary spintronic devices within a purely digital framework. This approach avoids the need for power-hungry analog-to-digital converters (ADCs) while leveraging the high density of Spin-Transfer-Torque RAM (STT-RAM).
Methodology
The GLM Neuron Model
At the core of SpinAPS is a neuron model based on the Generalized Linear Model (GLM). Unlike deterministic models, the GLM treats neuron outputs as probabilistic random processes. This stochasticity is implemented using a sigmoid activation function that determines a spiking probability, followed by a Bernoulli random variable generator.
First-to-Spike (FtS) Decoding
To minimize latency, SpinAPS employs a "first-to-spike" decoding rule. Instead of waiting for a full spike train to calculate a rate, the system makes a classification decision as soon as the first output neuron fires. This strategy significantly reduces the number of time steps required for inference.
Hardware Implementation
SpinAPS uses $n$-bit synaptic weights represented by $n$ binary STT-RAM devices. This digital representation ensures reliability and eliminates the complexity of analog memristors. The architecture is further optimized through:
- Piecewise Linear (PWL) Approximation: The sigmoid function is approximated using bit-shifts and additions to reduce logic complexity.
- Quantization: 5-bit quantization for weights and biases maintains accuracy while minimizing the hardware footprint.
- Tiled Mesh Architecture: A scalable design consisting of neuro-synaptic cores allows for expansion to support large-scale networks.
Experimental Results
The performance of SpinAPS was evaluated using the MNIST handwritten digit dataset and the Human Activity Recognition (HAR) dataset.
- Accuracy: SpinAPS achieved 93.5% accuracy on MNIST and 94.5% on HAR, results that are highly competitive with traditional Artificial Neural Networks (ANNs).
- Efficiency: The accelerator demonstrated a 4× improvement in GSOPS/W/mm² compared to equivalent SRAM-based designs. When scaled to a 32nm node, it reaches 850 GSOPS/W, making it approximately 20× more energy-efficient than high-end GPUs like the NVIDIA Tesla V100.
- Latency: On MNIST, the FtS rule allowed for correct classification of 75% of images within just 4 algorithmic time steps, showcasing the speed of the probabilistic approach.
Conclusion
SpinAPS demonstrates that combining spintronic memory with a probabilistic digital SNN framework can overcome the efficiency barriers of traditional AI hardware. By utilizing binary STT-RAM and first-to-spike decoding, the architecture provides a high-density, low-power alternative for edge AI applications. Future research will focus on integrating on-chip learning and testing the architecture on more complex, high-dimensional datasets.
Summary updated