Keeping Hot Chips Cool Thermal Management for Green Computing Yang Ge Professor Qinru Qiu.

Keeping Hot Chips Cool

Thermal Management for Green Computing

Yang Ge Professor Qinru Qiu

utline

• Background– Need for green computing– Adverse effects of high temperature– Thermal management techniques

• Ongoing project– Power and thermal management for single chip

cloud computer (SCC)

The need for green computing• Computers consume 3%

of US energy use– Saving 1% of energy of

data center is more than saving a power plant

• Each computer generates 1 ton of CO2 every year– Equivalent to the CO2

emission of a car driving a round trip between New York and Los Angeles

Power and Cost for Cooling Systems

• The energy dissipation for cooling system is high– Cooling fan power can reach

up to 51% of the overall server power budget

• The cooling cost is expensive in large data centers– The total cooling costs for

large data centers can run into tens of millions of dollars

Fans

CPU

Mem

OtherFans 51%

Mem 20%

CPU 24%

Other 6%

IBM P670 Server power breakdown

Adverse effects of high temperature to VLSI Chips

• Affects the system reliability and causes permanent device failure

• Doubles leakage power consumption every 9oC increase

• Requires to increase fan speed which could reduce fan life time

Thermal Management Techniques

Offline Techniques

Online Techniques

Temperature aware scheduling

Dynamic voltage frequency scaling

Temperature aware task migration

Ongoing Project

• Power and thermal management for single chip cloud computer (SCC)

• 24 tiles arranged in 6X4 arrays

• 2 CPUs on each tile

• A router associated with each tile

• 4 memory controllers go to on board memory

Overview of SCC Architecture

• SCC and MCPC communicates over PCIe bus

• MCPC runs Ubuntu 10.04 x64 and SW from Intel

• Load Linux image on each core

• read and modify SCC registers

• Load programs on the SCC cores.

Management Console PC (MCPC)

• 6 voltage domains• 24 Frequency

domains, one for each tile

• 2 temperature sensors on each tile

• Voltage and frequency can be changed separately on each domain

Power and Thermal Management

hank y u