(Processing Systems Lab)

Nov 10, 2017

Sung's work on memory-adaptive machine learning accelerators to be presented at DATE'18

Sung's paper on memory adaptive machine learning accelerators will be presented at the upcoming Design Automation and Test in Europe (DATE). The work exploits observations in both memory cell design and back-propagation to allow significant voltage-overscaling of SRAMs to reduce energy dissipation with marginal impact on accuracy. The work allows the system to reliably operate well below the SRAM Vmin, making the system "learn" around the kinds of hardware faults caused by memory supply overscaling.