Sushma Prasad

Abstract Title

Towards Ultra-scale energy efficiency
Achieving sustainable, high-performance computing at Google's ultra-scale requires an unyielding commitment to energy efficiency across every layer of the technology stack. This talk will delve into Google's holistic strategy, beginning with the foundational role of custom hardware design, where architecture and design optimizations drive significant power reductions. We'll explore the transformative impact of software-configurable optical interconnects, and how these help right size workload deployments. A critical component of this strategy is advanced liquid cooling, enabling denser server deployments and more efficient heat dissipation than traditional air-cooling methods. Further up the stack, custom compilers and intricate performance optimizations at the software level unlock additional energy savings by streamlining code execution and resource utilization. Finally, we'll discuss how Google wide fleet perf/W monitoring is unlocking new opportunities for energy efficiency. This comprehensive approach ensures that every joule of energy is maximized, underpinning Google's commitment to both technological leadership and environmental stewardship.

Biography

Sushma Honnavara-Prasad holds an MS in Electrical Engineering, specializing in Circuits and Systems, from the University of Texas, Dallas. Sushma began her professional journey working on OMAP platforms at Texas Instruments, followed by contributions to Smartphone SoCs at Broadcom. Notably, she was the chip lead for the inaugural product at Groq, an ML accelerator startup. Since joining Google, Sushma has played a pivotal role in the evolution of Tensor Processing Units (TPUs), engaging in various capacities including physical design, microarchitecture development, and hardware-software co-design. Her current endeavors are centered on full-stack machine learning energy efficiency, encompassing optimizations across data centers and individual chip architectures. Her core interests lie in the synergistic co-design of hardware and software for ML accelerators and pioneering energy-efficient computing solutions.

Contact

sushmahp@google.com