Collaborative Research: Feedback-Driven Resiliency for Near-Threshold Systems

Near-threshold voltage (NTV) operation has emerged as a promising strategy to improve energy-efficiency by reducing the supply voltage while exploiting parallelism to compensate for frequency loss. NTV processors combined with 'Turbo mode' can also offer peak performance, thereby making them suitable for a range of dynamic workload behaviors. Despite the promise, applications of these systems have remained largely elusive due to many issues. This research, involving VLSI circuits, computer architecture, and software systems, for the first time, addresses the reliability challenges of NTV/Turbo mode systems in the context of device aging and variability. Researchers will integrate distributed system monitors and controls to enable power-efficient fault resiliency for on-chip NTV/Turbo-mode cores. The work will lay the foundations for a feedback-directed optimization system that tunes the hardware's execution to meet application requirements by balancing performance, reliability and energy consumption.

Our society's productivity and advancements are intimately tied to the predictable and sustainable operation of the world's computing infrastructure. Unfortunately, this trend is beginning to falter due to increasing computer energy consumption. Aggressive energy reductions are tightly coupled with destabilizing computer system performance and longevity. The research will uncover observations that are necessary to overcome the problem in near-threshold voltage systems that hold great promise for further advanced computing. Using an approach spanning both the hardware and software layers, researchers will investigate and develop novel techniques to overcome the energy problem. Such a holistic and joint effort across the layers is the key for ensuring our society's long-term success based on advanced computing.