@conference {41923,
author={ Christopher Noel Hesse and Holger Eichelberger},
title={Benchmarking Neural Networks on Heterogeneous Hardware Resources},
publisher={ - CEUR-WS.org},
booktitle={CEUR-WS Proceedings of Symposium on Software Performance 2021 (SSP'21)},
year={2021},
url={http://ceur-ws.org/Vol-3043/short4.pdf},
abstract={<p>In recent years, artificial intelligence (AI) became a key enabling technology for many domains. To<br />achieve best performance, modern AI methods have high resource demands, e.g., GPU servers for the<br />training of neural networks. With the advent of further processor technologies, such as tensor processors<br />or re-wirable processors, AI methods may be executed in shorter time while even saving energy. For<br />many application domains such as autonomous driving or unmanned aerial vehicles, real-time constraints<br />mandate low end-to-end latencies in AI processing.<br />In this paper, we present a combined micro- and macro-benchmarking approach to analyze the<br />performance as well as the power demands of modern processor architectures using convolutional neural<br />networks as workload. We discuss tradeoffs among the different processor types and indicate issues and<br />challenges that arise when performing such benchmarks on heterogeneous hardware resources.<br />We show that FPGAs allow for an increase of 7x up to 45x in performance over high-end GPUs while<br />using only 10% of the power. In the consumer space, novel architectures such as the Apple M1 are able<br />to offer 3-5x better performance at 10-20% the power draw of current x86 CPU or GPU hardware.</p>}
}
