Instruction-aware Learning-based Timing Error Models through Significance-driven Approximations
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionThe adoption of aggressively down-scaled voltages along with worsening process variations render nanometer devices prone to timing errors that threaten system functionality. Recent studies on functional units focus on predicting these errors by exploiting machine learning-based error models. However, such models may be inaccurate when applied to complex pipelined designs, since they neglect important microarchitecture properties. In this paper, we propose a microarchitecture-aware ML model for timing error prediction that jointly considers various instruction types as well as all in-flight instructions in a pipeline. To circumvent the increased complexity, we utilize significance-driven approximations to improve inference time up to 4.66x, with less than 3% accuracy loss.