CAPA: Manufacturing Carbon Estimation for Advanced-Packaged Architectures

To meet growing compute demand, hyperscalers are rapidly deploying new data centre hardware, which embodies significant carbon. This aggressive growth of compute infrastructure jeopardizes their carbon reduction goals. While existing carbon tools estimate the manufacturing carbon footprint of integrated circuits (ICs), they do not reasonably model the high-performance advanced-packaged processors that dominate data centres and supercomputers. To bridge this gap, we propose a tool called CAPA, which models manufacturing Carbon for Advanced-Packaged Architectures. CAPA incorporates a binning yield model, high-bandwidth memory (HBM) carbon estimates, and support for complex architectures that use a mixture of integration techniques, such as 3.5D. By using CAPA to study widely used high-performance processors, we reveal opportunities for carbon savings through systematic binning and testing strategies, and we highlight HBM as a major carbon contributor. CAPA provides insights about the carbon footprint of cutting-edge architectures, opening new avenues for mitigating the environmental impact of the ever-growing computing industry.