Total Ionizing Dose Effect on Deep Neural Networks Implemented with Multi-Level RRAM Arrays
Sacane, Jack R.
MetadataShow full item record
This research work presents a methodology for simulating the effects of total ionizing dose (TID) radiation upon RRAM-based neural network accelerators. The experimental data on irradiating a 256×256 RRAM array test chip with Co-60 gamma rays up to a maximum TID of 1 Mrad (Si) were characterized with statistical methods in order to model the drift in RRAM cell conductance as a function of TID level. Multiple deep neural network (DNN) models were developed in the PyTorch framework in order to evaluate the effects of TID on DNNs implemented in hardware with similar RRAM memory technology and levels of radiation exposure. Using the statistical parameters discovered from the experimental TID data, weight changes were injected into the DNNs in order to simulate TID radiation effects and evaluate the resultant change of inference accuracy. Multiple simulations were conducted adhering to this methodology and the results pertaining to TID-induced inference accuracy degradation are discussed further in this work.