Reproduce “Memory Organization for Energy-Efficient Learning and Inference in Digital Neuromorphic Accelerators” on OpenResearch

About this paper

Clemens JS Schaefer, Patrick Faley, Emre O Neftci, Siddharth Joshi

Researchers from the University of Notre Dame and UC Irvine developed and evaluated various memory organization schemes for digital neuromorphic accelerators, including a novel functional encoding for convolutional layers. The study directly measured energy costs from synthesized hardware, demonstrating a 58% reduction in backward pass energy for convolutional layers using functional encoding and a 1.37x energy efficiency improvement for sparse networks with pointer-based bitmap (PB-BMP) at a 4% accuracy trade-off.

Key insights

• Direct energy measurements revealed that memory size can have a substantial and often dominant impact on total energy consumption, challenging previous assumptions focused primarily on memory access counts.
• The most energetically favorable memory organization scheme depends critically on network characteristics such as sparsity, layer type (fully connected vs. convolutional), and desired weight quantization level.
• Functional encoding provides considerable energy savings for the backward pass in convolutional layers, highlighting its potential for efficient on-chip learning in SNNs with structured connectivity.

Summary by alphaXiv.