The Effect of Heavy-Ion Single Events on the Accuracy of SONOS Analog In-Memory Computing Accelerators

Abstract

Silicon-oxide-nitride-oxide–silicon (SONOS) charge-trap memory is a promising nonvolatile memory (NVM) device for analog in-memory computing (AIMC) applications. The sensitivity of 40-nm SONOS analog states in response to single-event effects (SEEs) was experimentally assessed by exposing an SONOS array to heavy ions with the linear energy transfers (LETs) of 28.6, 44.9, and 59.5 MeV/(mg/cm2). A statistical model of SEE was developed using specific criteria to distinguish them from effects unrelated to radiation, such as device-to-device programming variation, read noise, and drift. This model was used to simulate the effects of heavy ions on an AIMC accelerator processing standard deep neural networks (DNNs) running inference on common datasets, including ImageNet. Simulated inference accuracy decreases with increasing LET and fluence. At the highest LET of 59.5 MeV/(mg/cm2), an accuracy degradation of 5% on ImageNet was predicted at a fluence of $3\times 10^{6}$ ions/cm2, indicating a sufficient performance for most space applications.