How Design Choices Shape Sustainable Computing: A Hardware Accelerator Perspective

Data centers are experiencing rapidly rising energy consumption and associated greenhouse-gas emissions. Today, GPUs are the dominant accelerator in many data centers, yet they are not necessarily the most efficient choice for every application domain because they remain relatively general-purpose. In this talk, we examine how accelerator design choices shape sustainable computing outcomes through case studies in genomics and AI. First, we present an FPGA-based, parameterized accelerator family tailored to a representative genomics workload and designed to expose key architectural and system-level trade-offs. A hardware–software co-design approach reveals optimal operating points in terms of time-to-solution and energy-to-solution, outperforming GPU baselines across a range of problem sizes and deployment constraints. Second, we explore emerging trends in AI acceleration and energy efficiency, and how they influence total emissions and costs depending on local carbon intensity and electricity prices. Across both use cases, we find that the economically optimal configuration can diverge from the emissions-minimizing configuration, highlighting a gap between cost-driven procurement and climate-friendly operation in data centers. We conclude with practical co-design guidance to align performance goals with sustainable computing targets.