|M.Sc Student||Zobel Shmuel|
|Subject||Power Performance Tradeoffs in Graphics/GPGPU Based Systems|
|Department||Department of Electrical Engineering||Supervisors||Professor Avi Mendelson|
|Professor Emeritus Avinoam Kolodny|
|Full Thesis text|
The increased performance of graphics processors and their unique design led the industry to seek additional usage models besides graphics for these processors. Led by Nvidia, many GPGPU (General Purpose Graphics Processor Unit) applications started to show up. Many works have been published dedicated to increasing the performance of the applications both from the hardware and software sides, but very small attention was given to the tradeoffs between performance, power dissipation and maximal temperature. Power issues are critical as some graphics cards failed or had to be recalled due to too high power consumption or excessive operating temperatures and due to the fact that power can be always traded with performance with power if the maximum power or maximal thermal allowed were not reached.
In this research we concentrated on the power, temperature and performance tradeoffs. The work is experimental and based on real measurements of several Nvidia systems. We found that, in general, increasing the parallelism; e.g., having more threads to execute, results in improving the performance. On the other hand it also increases the power. Taking power and thermal consideration in account shows that the gain from deeper parallelism saturates or may even degrade. For example in matrix multiplication on GTX285 increasing number of blocks from 1 to 1024 shorted the execution time from 237s to 5.2s but the farther increase to 262144 blocks helped to reduce to 4.8s, and the energy optimization was at 16384.
Optimizing the system to achieve best energy consumption, yields that amount of parallelism the application exposes, may be lower than reported and what one could expect. Concentrating only at active power (assuming leakage power is zeroed by some magic way like running at low voltage or low temperature or new transistors technology) the tradeoffs shows that the optimal number of cores is even lower than before.
Temperature has direct correlation with power consumption and so we can assume that the higher the power the hotter the GPU becomes. Thus, computer architectures depend on thermal density and not only on total thermal of the chip. In addition, when measuring the temperature using the thermal sensors, we found that some applications showed degradation of local temperature although total power increases. For example in the reduction algorithm, execution time reduced from 1s to 0.155s with the cost of increase of total power from 92W to 145W but Max temperature was reduced from 58˚C to 51˚C.