A Performance and Power Comparison of Contemporary GPGPU Architectures

General-purpose Graphics Processing Units (GP-GPUs) have been extensively studied over the last fifteen years. GP-GPUs execute general purpose applications on a graphics processing instead of the traditional graphics applications. GP-GPUs employ a single instruction multiple thread (SIMT) execution wherein there are multiple processors and thousands of light-weight threads. In this article we compare the performance and power consumption of a variety of graphics processors. Specifically we compare Fermi, Kepler, Pascal, Turing, and Volta graphics processors. These are processors that were released from more than a decade back (Fermi) to the most recent processors (Volta). We observe that for a good number of applications Volta gives the best performance. Surprisingly, for few applications Pascal gives the best performance. We root-cause the reason for such a behavior and observe that the memory-level-parallelism (MLP) is the reason with Pascal reporting a higher MLP than Volta. We also observe that the power consumption mirrors the performance with identical behavior. We conclude that the performance of a GPU has many variables and only moving to a newer generation of GPU architecture will not always give the best performance.