A Novel Approach to Hardware-Software Co-Design for Power-Efficient AI Systems
Karthik Wali · Journal of Artificial Intelligence Machine Learning and Data Science · 2022
The increasing number of AI applications, various usages, and diversified models have prompted the need for a more advanced and reusable hardware platform.Nonetheless, the large number of computations that AI algorithms require is a crucial challenge in edge computing, embedded systems, and battery-operated devices regarding energy consumption.Consequently, in this paper, a hardware-software co-design framework is presented to enhance the performance of the AI system and, at the same time, reduce its energy consumption.With the help of our combined approach, which includes HW-NAS, DVFS, and low-bandwidth AI models forming as well as compiler-level pruning, quantization, and model compression, our methodology achieves the best trade-off in terms of performance, accuracy, and energy.To this end, our design focuses on Field-Programmable Gate Arrays (FPGAs) and System-On-Chips (SoCs), obtaining portable and transferable solutions for various AI tasks.Promising experiments also show up to 50% energy saving with a very less hit in performance on the different AI models.Therefore, this research provides a solution for improving the efficiency of AI power consumption and enhancing real-time and other embedded applications.