Accurate Measurements and Precise Modeling of Power Dissipation of CUDA Kernels toward Power Optimized High Performance CPU-GPU Computing

Abstract

Abstract-Power dissipation is one of the most imminent limitation factors influencing the development of High Performance Computing (HPC). Toward power-efficient HPC on CPU-GPU hybrid platform, we are investigating software methodologies to achieve optimized power utilization by algorithm design and programming technique. In this paper we discuss power measurements of GPU, propose a method of automatic extraction of power data of CUDA kernels from long measurement sequence, and execute an exactitude and effective power analysis on CUDA kernels. By using the proposed method above, we measured a sample kernel that performs single precision floating point additions on GeForce 8800 GTS. Our results suggest that the power consumption by a nonworking thread in underoccupied half-warp is 71% of the power consumed by a working thread.

Power versus performance tradeoffs of GPU-accelerated backprojection-based synthetic aperture radar image formation

“…17 Their work illustrates how to measure the power of a GPU device by multiplying the voltage and the current measurements captured by external probing equipment. Specifically, they 1) attach voltage and current probes to each power input, i.e., voltage rail, of a GPU; 2) multiply the current and voltage values to compute the power for each rail; and 3) add the power for each rail to get the total power usage of the GPU.…”

GPU Acceleration on Embedded Devices. A Power Consumption Approach

“…However increasing the number of processing units also consumes more energy, thus, it is important to know when to enable or disable processing units. On [9] [10] it is shown information to optimize kernels in order to save power consumption. To get the performance of these systems a specialized benchmark can be used [11].…”

Power profiling and optimization for heterogeneous multi-core systems

“…There are studies in power modelling for FPGAs and GPUs and various models have been proposed [1][2][3][4][5][6]. Some of these studies provide analytical treatment of device power behaviour.…”

“…In 2009, power measurement of the major components inside a GPU platform including the processing cores and the memory hierarchy was reported [7]. In the same year, more comprehensive experiments are performed [6]. Using CUBLAS and ATLAS as the workload in the experiment, realtime measured values are identified by inserting special workload as time markers.…”