Constructing Application-level GPU Error Rate Model with Neutron Irradiation Experiment

This work explores GPU application factors that dominantly characterize SDC and DUE cross sections obtained by neutron irradiation experiments. We evaluate correlation coefficients between the cross sections and application metrics obtained from profiler outputs and fault injection results. Experimental results with six applications show that “warps per block” is a primary factor for both SDC and DUE. The DUE cross section modeling is improved by the secondary factor of L2 hit rate.