Analysis of the Analytical Performance Models for Gpus and Extracting the Underlying Pipeline Model

Abstract

Analytical models for the performance estimation of GPUs can be categorized into five main classes, as follows: Roofline, Volkov, MWP-CWP, WFG model and GPUMech. Volkov’s model and the equivalent Transit model provide an upper bound for the performance with the occupancy roofline. The MWP-CWP, the WFG model and GPUMech take additional performance-limiting factors into account. We show that each model can be derived from a pipeline analogy that models each GPU subsystem as an abstract pipeline. We present the novel Pipeline model as the underlying model that underpins the analytical equations of these models and makes their simplifying assumptions explicit. Rather than relying on equations, the Pipeline model is then used to simulate the behavior of kernel executions on an abstract level based on the same hardware parameters as the analytical models. The simplicity of the model and relying on simulation mean that this approach needs less assumptions, is more comprehensive and is more flexible. More performance aspects can be taken into consideration. The different models are compared and evaluated empirically with 14 kernels of the Rodinia benchmark suite. The Pipeline model gives an average MAPE of 23, while the average MAPE values of the other models lie between 27 and 136.

Keywords: GPU Computing, GPU architecture, Analytical Performance Models