Characterization and Machine Learning Classification of AI and PC Workloads

Abstract

To better design AI processors, it is critical to characterize artificial intelligence (AI) workloads and contrast them to normal personal computer (PC) workloads. In this work, we profiled the AIBench and PassMark PerformanceTest benchmarks with the Intel oneAPI VTune Profiler on a multi-core computer. We captured and contrasted the various CPU and platform metrics and event counts for these two distinct benchmarks. Using the Orange 3.0 data mining tool, and based on the captured profile metrics and event counts, we then trained and tested 9 machine learning (ML) models to classify the CPIs and elapsed times of the various tests of these two benchmarks, including inference and training tests in AIBench, and CPU, memory, graphics, and disk tests in PassMark. The linear regression machine learning model emerged as the best clocks per instruction (CPI) classifier, while the neural network model with 4 hidden layers was the best elapsed time classifier. This machine learning classification can help in predicting the CPI and elapsed time and distinguish between AI and standard PC workloads based on the profiled application(s) and captured profile metrics and event counts. The stressed computer units identified by this detailed profiling work and exercised by the benchmark tests can also guide future AI processor design improvements.

Keywords

• AI workloads
• Tensorflow
• PassMark PerformanceTest
• AIBench
• workload characterization
• event counts