Performance Analysis of AI workload on Intel hardware platform

Abstract

The increased popularity of Deep Neural Network (DNN) has led to a good amount of research in getting the best hardware configuration to achieve the optimum throughput from server processors running an end to end DNN topology. The pri- mary focus here is to do the inference which is to know how efficiently the processor can execute the already trained models for the Text To Speech (TTS) topology. This thesis explores the benchmarking performed on Intel's server platform with DeepMind's Tacotron-2 topology as a workload. Benchmarking consists of running the topology using baseline and Intel optimized Tensor ow framework. The timeline analysis and system performance analysis helps in understanding the behavior of the topology on the server platform. The processor's system performance is measured using Intel's VTune amplifier tool. In order to understand the impact of workload on the server processor, this thesis focuses on the key parameters like Processing time, Memory bandwidth utilization and Memory latency, Hotspot analysis, and effective Central Processing Unit (CPU) utilization. Along with the above param- eters, timeline and baseline pro ling on the workload is also done. As an outcome of this benchmarking, several observations and recommendations are proposed for further research and improvements that can be done in the area of Silicon architec- ture, Software frameworks, and Kernels. The results highlight the key observations and the performance graphs and comparison using baseline and Intel Optimized framework for the tacotron module.