Device-aware inference operations in SONOS nonvolatile memory arrays
Authors:
Christopher H. Bennett,
T. Patrick Xiao,
Ryan Dellana,
Vineet Agrawal,
Ben Feinberg,
Venkatraman Prabhakar,
Krishnaswamy Ramkumar,
Long Hinh,
Swatilekha Saha,
Vijay Raghavan,
Ramesh Chettuvetty,
Sapan Agarwal,
Matthew J. Marinella
Abstract:
Non-volatile memory arrays can deploy pre-trained neural network models for edge inference. However, these systems are affected by device-level noise and retention issues. Here, we examine damage caused by these effects, introduce a mitigation strategy, and demonstrate its use in fabricated array of SONOS (Silicon-Oxide-Nitride-Oxide-Silicon) devices. On MNIST, fashion-MNIST, and CIFAR-10 tasks, o…
▽ More
Non-volatile memory arrays can deploy pre-trained neural network models for edge inference. However, these systems are affected by device-level noise and retention issues. Here, we examine damage caused by these effects, introduce a mitigation strategy, and demonstrate its use in fabricated array of SONOS (Silicon-Oxide-Nitride-Oxide-Silicon) devices. On MNIST, fashion-MNIST, and CIFAR-10 tasks, our approach increases resilience to synaptic noise and drift. We also show strong performance can be realized with ADCs of 5-8 bits precision.
△ Less
Submitted 2 April, 2020;
originally announced April 2020.