Carbon-Aware Design of DNN Accelerators: Bridging Performance and Sustainability

The rise of Machine Learning (ML) and the wide use of Deep Neural Networks (DNNs) have led to the development of specialized DNN accelerators, aimed at improving the computing capabilities of ML systems. While these accelerators have been beneficial in enhancing computational efficiency, their design presents significant challenges due to the complexity of hardware configurations. Additionally, it is critical to consider and address the environmental implications and challenges. Unfortunately, these aspects have often been neglected during the design of ML systems. The carbon footprint, both operational and embodied, of these accelerators is a growing concern, with the latter becoming increasingly significant. This paper presents a framework for designing DNN accelerators with an emphasis on embodied carbon footprint. Using a genetic algorithm, we address the balance between performance and sustainability, focusing on reducing the embodied carbon footprint. Experimental results on different types of DNNs show that our method, by exploiting the properties of on-chip logic and memory, can generate DNN accelerators with considerably less embodied carbon and with a negligible performance overhead.