SparkXD: A Framework for Resilient and Energy-Efficient Spiking Neural Network Inference using Approximate DRAM
Hosted in Virtual Platform
Emerging Models of Computation
DescriptionThe DRAM accesses dominate the energy consumption in processing spiking neural networks (SNNs). State-of-the-art in SNN-systems do not optimize the DRAM energy-per-access, thereby hindering achieving high energy-efficiency. To minimize the DRAM energy, a key knob is to reduce the DRAM supply voltage, but this may lead to DRAM errors (i.e., the so-called approximate DRAM). Towards this, we propose SparkXD, a novel framework that employs 1) error-aware SNN training; 2) SNN error-tolerance analysis to find the tolerable error rates; and 3) error-aware DRAM mapping. For 1% tolerable accuracy loss, SparkXD reduces the DRAM energy by 39.46% on average.