MONETA: A Processing-In-Memory-Based Hardware Platform for the Hybrid Convolutional Spiking Neural Network With...

Area of Science:

• Artificial Intelligence
• Computer Architecture
• Neuroscience

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation inspired by biological brains.
• Processing-in-memory (PIM) architectures aim to reduce data movement bottlenecks in deep learning.
• Hybrid SNNs combine supervised and unsupervised learning for enhanced capabilities.

Purpose of the Study:

• To develop and evaluate MONETA, a PIM-based hardware platform for accelerating hybrid convolutional SNNs.
• To enable on-chip, on-line training and inference for SNNs.
• To assess the performance and accuracy of MONETA on benchmark datasets.

Main Methods:

• Designed MONETA using 8T SRAM-based PIM cores for vector matrix multiplication (VMM) and spike-time-dependent plasticity (STDP) weight updates.
• Implemented an SNN-focused data flow to minimize data movement and maintain learning accuracy.
• Evaluated MONETA with 4-bit input and 8-bit weight precision on the CIFAR-10 dataset.

Main Results:

• MONETA achieved competitive accuracy on CIFAR-10, with only a 1.63% drop compared to software-based STDP for convolutional SNNs (ConvSNNs).
• The hybrid SNN architecture accelerated by MONETA showed a 10.84% accuracy improvement over STDP-only training and 1.4% over backpropagation-based ConvSNNs.
• Physical design in 65nm CMOS demonstrated significant power efficiencies: 18.69 TOPS/W (inference), 7.25 TOPS/W (learning), and 10.41 TOPS/W (hybrid learning).

Conclusions:

• MONETA provides an efficient PIM hardware solution for accelerating hybrid SNNs, supporting both inference and on-chip learning.
• The proposed data flow and architecture enable accurate and efficient on-line training and inference for SNNs.
• The high power efficiency and competitive accuracy highlight the potential of MONETA for edge AI applications.