Collaborative Research: DESC: Type I: Transforming Integrated Electro-Photonic Fabrics for Light Speed Communication and Computation

Overview Carbon emissions from cloud datacenters account for 0.6% of global carbon emissions today. This number is projected to increase dramatically over the next decade due to the explosion of everyday consumer electronic devices and systems empowered by cloud datacenters. Due to the exacerbating impact of carbon emissions on global climate change, researchers have recognized the need for realizing carbon-efficient and sustainable datacenters. To that end, several prior efforts have focused on minimizing the operational (use-phase) energy consumption and related carbon costs in datacenters. However, sustainable design is not merely about minimizing the use-phase carbon costs. It is also crucial to minimize or reduce embodied carbon emissions from infrastructure development activities and the manufacturing of computing hardware. Besides, sustainable design should account for reliable lifespan, replacement rate, and utilization efficiency of critical computing hardware resources of datacenters. The vital hardware resources in modern datacenters include DRAM-based architectures for data provisioning and processing. A modern datacenter typically deploys DRAM-based architectures scattered across computing servers or disaggregated in massive-capacity pools interconnected using the Compute Express Link (CXL) standard. These DRAM- based architectures, irrespective of their deployment topology, often present daunting performance and carbon bottlenecks; therefore, they play a crucial role in determining the sustainability of datacenters. Unfortunately, how to improve DRAM-based architectures and their usage for sustainable data provisioning and processing in datacenters remains unknown. To overcome this shortcoming, this pioneering project will present well-reasoned and synergistic research activities that will focus on the overall carbon modeling of DRAM-based architectures, analysis of their carbon bottlenecks and trade-offs, minimization of hardware idling and carbon wastage, and prolongation of carbon-efficient lifespan. This project will also create opportunities for industrial partnerships, local outreach, diversity promotion, and curriculum innovation. Keywords: DRAM; Carbon Efficiency; Datacenters; Aging Effects; Reliable Lifespan; Embodied Carbon Intellectual Merit This project will employ the principles of heterogeneity, hardware polymorphism, and recycling to design carbon-efficient DRAM-based architectures to transform the sustainability, performance, resource utilization, and reliable lifetime of data provisioning and processing in datacenters. The overarching goal of this project will be to dramatically reduce the embodied and use-phase carbon footprints and prolong the reliable lifespans of DRAM-based architectures. This project will result in (1) a framework, developed through proposed collaborations with industry partners, that factors in the critical impacts of device technology options, microarchitecture organizations, hardware resource mapping and re-appropriation techniques, and various aging effects for modeling the embodied and use-phase carbon footprints of DRAM-based architectures, (2) an extensive simulation framework for workload-driven evaluation, validation, and comparison of DRAM-based architectures, focusing on various important metrics for carbon efficiency, performance, lifetime, and resource utilization, (3) proactive and reactive techniques to prolong the reliable and carbon-efficient lifespan of DRAM-based architectures, (4) methods to repurpose DRAM fabrics to minimize resource idle-time and carbon wastage, and (5) strategies for cross-layer design and co-location of processing and memory arrays to achieve simultaneous reductions in embodied and use- phase carbon footprints of DRAM-based architectures for data provisioning and processing. Broader Impacts The outcomes of this research project will ignite a wide spectrum of future research efforts, which will inform U.S.-based stakeholders of sustainable practices to follow for designing DRAM-based architectures for data provisioning and processing in datacenters. This will help the U.S. out-compete global competitors to emerge as the pioneer in sustainable datacenter design using DRAM-based architectures. Besides, this project will also create novel educational materials using the newly developed tutorial videos to give students a more tangible, hands-on approach to learning the fundamentals of the structure and operation of DRAM-based architectures and principles of sustainable hardware design. These materials will be employed to (i) innovate curricula, (ii) promote community outreach, (iii) engage undergraduate and graduate students in supervised research, and (iv) promote diversity by training STEM teachers at local middle schools that primarily serve underrepresented groups. These educational activities will help enlarge our nation’s workforce in DRAM-based hardware design and sustainable computing.