SPECnet: Predicting SPEC Scores using Deep Learning

Abstract

In this work we show how to build a deep neural network (DNN) to predict SPEC® scores - called the SPECnet. More than ten years have passed since the introduction of the SPEC CPU2006 suite (retired in January 2018) and thousands of submissions are available for CPU2006 integer and floating point benchmarks. We build a DNN which inputs hardware and software features from these submissions and is subsequently trained on the corresponding reported SPEC scores. We then use the trained DNN to predict scores for upcoming machine configurations. We achieve 5%-7% training and dev/test errors pointing to pretty high accuracy rates (93%-95%) for prediction. Such a prediction rate is very comparable to expected human-level accuracy of 97%-98% achieved via careful performance modelling of the core and un-core system components. In addition to the CPU2006 suite, we also apply SPECnet to SPEComp2012 and SPECjbb2015. Though the reported submissions for these benchmark suites number in hundreds only, we show that such a DNN is able to predict for these benchmarks reasonably well (~85% accuracy) too. Our SPECnet implementation uses state-of-the-art Tensorflow infrastructure and is extremely flexible and extensible.