A64FX Enables Engine Decarbonization Using Deep Learning

Decarbonization of transportation requires new deep learning models to enable improved engine control. Research and development must also be done in a computationally efficient manner, so of interest is to understand the applicability of high performance computing resources to be used for training machine learning models and to compare both the power consumption and temporal performance of new A64FX architectures to x86 architectures. This work details the development of a Multilayer Perceptron (MLP) model using the Fujitsu A64FX processor for predicting in-cylinder pressure of internal combustion engines, a critical performance parameter to develop pathways to decarbonized engine controls. A scaling analysis of up to 160 compute nodes demonstrates continuously improving performance with increasing number of nodes. A comparison of performance and power consumption between A64FX and Intel's x86 Sapphire Rapids (SPR) is also included and shows that up to 30 parallel nodes the power efficiency of A64FX is lower, but that its energy consumption is constant as opposed to SPR which increases linearly as node count increases.