SHF: Medium: Collaborative Research: Predictive Modeling for Next-generation Heterogeneous System Design

University Of California-Riverside, Riverside CA

Investigators

Abstract

With semiconductor scaling reaching physical limits, performance and power consumption are ever more critical aspects in the design of emerging computer systems. Fast and accurate design models and tools are critical to support future computer system designers in evaluating design options before they can be built. Traditional simulation-based or analytical models are often too slow or inaccurate to effectively support design processes. This project instead develops novel machine learning-based, predictive methodologies to rapidly estimate the performance and power consumption of future generation products at early design stages using observations obtained on commercially available silicon today. Such techniques will allow efficient design cycles ensuring that the next-generation computing infrastructure meets the needs and expectations of consumers and continues to meet them over the product lifecycle. Along with research activities, course material on predictive modeling will be integrated into the university courses taught by the investigators, technology will be transferred to industrial partners through training and tutorials, and tools and models developed in this project will be released as open source software. In addition to training of graduate students, emphasis will be paid to undergraduate student training, towards including federally recognized under-represented groups, training of STEM teachers, and to run summer code camps to increase access for middle school and high school students. This project specifically investigates use of advanced machine learning techniques for prediction of power and performance of any machine based on hardware-dependent and independent application characteristics obtained by running on any existing other machine, focusing on large-scale data center and accelerator technologies, namely multi-core CPUs, GPUs and FPGAs. Specific research tasks include the investigation of: (1) fast and accurate models for system designers and system programmers to perform rapid, early hardware and software design space exploration; (2) fast online prediction models that can be integrated into modern operating systems and virtual machine; and (3) fast yet accurate model training procedures that can create new predictive models while applications run. This research is expected to also allow semiconductor companies to better understand the scenarios under which predictive modeling is sufficiently accurate to be deployed during an industrial design process. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.