Carbon-Efficient Design Optimization for Computing Systems

The world's push toward an environmentally sustainable society is highly dependent on the semiconductor industry, due to carbon footprints of global-scale sources such as computing systems for virtual and extended reality applications (VR and XR). Despite previous carbon modeling efforts for such computing systems, there lacks a wide range of design tools to optimize total life cycle carbon footprint (during manufacturing and also during day-to-day operation), while meeting application-level constraints (power, performance, area). To address this need, we have developed a carbon-aware design framework that optimizes carbon efficiency of computing systems-quantified by metrics such as total Carbon Delay Product (tCDP: the product of total carbon and total application execution time)-while also identifying key design parameters for improving carbon efficiency. As a case study, we demonstrate the effectiveness of our framework to improve tCDP of hardware accelerators for artificial intelligence (AI) and XR applications. We show: (1) optimizing for carbon efficiency (tCDP) instead of energy efficiency (Energy-Delay Product or EDP) improves carbon efficiency by up to 6.9x-i.e., optimizing for EDP is insufficient; (2) for multi-core CPUs inside production VR headsets, optimizing number of cores (from 8 to 4) improves tCDP by 1.25x (over their entire lifetime); (3) leveraging an advanced three-dimensional integration (3D) technique (3D stacking of separately-fabricated logic and memory chips) can improve tCDP by 6.9x vs. conventional systems (no 3D stacking).