Personal tools
Home Events Events Archive 2012 Effects of Cooling on the Overall Power Consumption of Processors

Effects of Cooling on the Overall Power Consumption of Processors

— filed under:

  • PhD Defenses
When Mar 09, 2012
from 10:00 AM to 12:00 PM
Where ENGR. IV Bldg. Tesla Room 53-125
Add event to calendar vCal

Won Ho Park

Advisor: C.K. Ken Yang



To address the power and thermal problems of high-performance computing systems, the possibility of using advance cooling systems is explored to realize overall system power improvement that includes the cost of cooling power. Realistic system-level modeling  that includes the electronic and refrigeration systems will expedite the analysis of the optimal operating temperature points, the amount of total power reduction at reduced temperatures, and the dependence of the refrigerated performance on the power profile of the electronics. Highly-efficient miniature-scale refrigeration system for electronic cooling is developed to experimentally demonstrate the amount of total power saving for processors at various operating conditions. A processor that dissipates 175.4W of maximum power with 30% electronic leakage power operating at 97°C is cooled using our refrigeration system. Measurements show that with a minimum refrigeration COP of 2.7, the processor operates with junction temperature <40°C and offers a 25% total system power reduction over the non-refrigerated design. Not only the measurement results validate our system-level model, but this experiment is the first demonstration of active cooling that lead to reduced total wall power. Furthermore, a model that captures different relations and parameters of multi-core processor and the refrigeration system is constructed based on the measured results. This model is used to present an energy-efficient workload scheduling that optimizes power efficiency under the actively cooled environment that can potentially be applied to large scale multi-core, multi-processor computing environment. Finally the proposed methodology is combined with the G/G/m-models to reduce both total power and response time degradation while meeting target SLA requirements.



Won Ho Park is currently a Ph.D. candidate at UCLA.

Document Actions