Zoom: https://illinois.zoom.us/j/88181700267?pwd=sTbTte6jTSBllXABlNlcxNbCqnbZKb.1
Abstract:
Modern applications, including Generative AI, are pushing computing to its limits, demanding more efficient hardware and software than ever before. Traditionally, optimizations have been siloed: focusing on either applications, software, or hardware in isolation. However, to sustainably scale performance and energy efficiency of AI-based applications, we must rethink system design holistically, embracing co-design across the stack to unlock new levels of efficiency. In this talk, I will present two of our works that exemplify this co-design philosophy. First, I will introduce Prodigy, a hardware-software co-design that accelerates classical workloads like graph analytics and scientific computing by rethinking the interface between software and hardware. Next, I will present MoDM, which applies application-system co-design to diffusion model inference, enabling dynamic trade-offs between performance and image quality using a mixture of models. Together, these works showcase how reimagining system design across abstraction layers can drive breakthroughs in both traditional and AI-driven computing. I will conclude by outlining my vision for the next frontier in AI-era computing: leveraging co-design to push beyond the limits of today's systems.
Bio:
Nishil Talati is an Assistant Research Scientist in the CSE department at the University of Michigan, where he also earned his PhD. His research focuses on computer architecture and systems software design to enhance the efficiency of modern data-driven applications. Nishil’s work has been featured in top-tier venues such as ISCA, MICRO, HPCA, ASPLOS, and VLDB, and has made a significant impact through industry tech-transfers, inspiring follow-up research, and earning several accolades. These include the IEEE computing’s top 30 early career professional award, HPCA Best Paper Award, honorable best paper mentions at DATE 2023 and IISWC 2023, recognition as a 2023 ProQuest Distinguished Dissertation Award finalist, and the Best Faculty Research Pitch Award at MIDAS event in 2023.
Faculty Host: Josep Torrellas
Meeting ID: 881 8170 0267
Password: csillinois