Exploiting GPU Undervoltage to Improve the Energy Efficiency of Deep Learning Applications

Rafael Gil

Abstract:—The success of deep learning applications, within machine learning and artificial intelligence, is pushing further this area’s development. However, the increasing performance and accuracy needs are usually met with higher computational requirements, whose efficiency is, more often than not, disregarded. General Purpose Graphics Processing Units (GPGPUs), being the state-of-the-art accelerators for these applications, play a significant role in making deep learning models widely available. However, the large power consumption increases operational costs and eschews resource-constrained environments from using such devices. To mitigate this problem, the present work proposes an approach to study the potential energy savings of reducing the supply voltage of those devices, using an AMD Radeon Vega Frontier Edition GPGPU. This endeavor is first applied to synthetic benchmarks to characterize the device’s voltage guardband and then to current deep learning models to provide an insight into their behavior under minimum supply voltage. Results show deep learning models can achieve energy savings of up to 24 . 79% (average of 15 . 35% ) and still guarantee their initial accuracy. Nonetheless, the energy savings can be further increased up to 30 . 16% (average of 18 . 37% ) at the expense of the model’s accuracy. Deep learning applications experienced an accuracy droop up to 61 . 52% (average of 10 . 61% ) when working at near failure supply voltage.

Computer Science,Engineering