Loading...

Improving GPGPU Performance Through Efficient use of Memory Controllers

Bakhishi, Mohammad Hazhir | 2016

816 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 49057 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sarbazi Azad, Hamid
  7. Abstract:
  8. Appearance of CUDA architecture results in GPU introduction as a suitable platform for parallel processing. Massive usage of GPGPUs in various applications forces the manufacturers to produce this processor with different configurations based on their application demands. But always the design approach observes a ratio between processing capability and bandwidth of GPGPUs. In this way mostly in all GPGPU series the bandwidth increases with growth of GPGPU processing power. At first glance this ratio seems reasonable because more process needs more data. However this approach does not pay attention to the behavior of a wide range of workloads which do not need such a bandwidth. Mentioned reason cause to low bandwidth utilization of GPGPU during execution of these workloads while utilization of execution units and on-chip resources is still high. Therefore maintenance of this ratio between processing capability and bandwidth capacity is not necessary for all of existing workloads. In this project we first introduce novel design approach for GPGPUs specifically for aforementioned workloads which attempts to trade the amount of off-chip bandwidth for more processing and on-chip resources. Bandwidth reduction results in removing extra memory controllers. Our investigations show that the occupied area by the memory controllers is very large. Therefore in the second step we attempt to utilize the area obtained from bandwidth reduction to improve GPGPU performance. Final simulation results show that the performance and energy of suggested architecture is improved 23% and 11% respectively during execution the mentioned workloads while performance under other workloads which utilized bandwidth efficiently does not decrease for more than 4% on average
  9. Keywords:
  10. Bandwidth ; Performance Improvement ; Energy ; Graphics Procssing Unit (GPU) ; Memory Controller ; Process Capability ; Utilizability

 Digital Object List

 Bookmark

No TOC