Conventional microprocessors are designed primarily for sustained performance; they can operate at near-peak performance essentially indefinitely until their energy source is exhausted. In battery and cooling constrained environments such as mobile devices, sustained power (and, consequently, peak performance) must be limited to at most a few watts so that the device can dissipate heat using only passive convection. However, many interactive mobile applications (such as handwriting/speech recognition or image-based search) instead call for bursts of intense computation in response to user input, creating the need for a new ultra-responsive operating regime: rather than limit peak power assuming sustained operation, systems should instead exploit heat storage to enable brief computation bursts that greatly exceed sustainable thermal limits without overheating. This project investigates an approach called "computational sprinting", the central essence of which is to compute at unsustainable rates but only briefly so that temperatures do not reach unsafe levels.

The goal of this project is to address the architectural, thermal, electrical, and software barriers to ultra-responsiveness via computational sprinting. In particular, the project explores: (1) architectures and memory systems that sprint via parallelism (activating tens of reserve functional units/cores) and voltage boosting (overdriving cores for single-thread performance) while facilitating fast burst onsets; (2) thermal designs that improve thermal response behavior to enable longer and more intense sprints; (3) mobile-optimized electrical designs that provide stable supply voltages despite current surges of an order-of-magnitude or more; and (4) software mechanisms that explicitly manage limited thermal budgets and anticipate and stage data needed during the sprint. The project includes the fabrication of an experimental testbed to approximate the computational, thermal and electrical capabilities of a future many-core mobile device and to provide practical experience with sprinting. Project impacts include (1) developing and advancing techniques for improving the responsiveness of mobile devices and (2) integrating the discoveries into a new cross-departmental course on computer system design.

Agency
National Science Foundation (NSF)
Institute
Division of Computer and Communication Foundations (CCF)
Application #
1161505
Program Officer
tao li
Project Start
Project End
Budget Start
2012-05-01
Budget End
2017-04-30
Support Year
Fiscal Year
2011
Total Cost
$560,000
Indirect Cost
Name
Regents of the University of Michigan - Ann Arbor
Department
Type
DUNS #
City
Ann Arbor
State
MI
Country
United States
Zip Code
48109