Analytical Model Templates for CPU and GPU Power and Performance
Purpose and Function
The purpose of the tool is to demonstrate (1) the integrated GPU and CPU performance and, (2) power consumption models. For demonstrating the performance model, the tool provides a differential frame time prediction model template. The template contains features used to predict the frame time and parameters that are learned at run time. The tool also implements an adaptive filter based on recursive least squares estimation (RLS) that minimizes the frame time prediction error at run time. The tool also contains several workload traces collected from Intel Baytrail platform. Given the input traces, the tool predicts the frame time, compares it to the measured data, and finally reports the error in frame time prediction as well as F-statistic for each workload. Since the input traces are obtained from real platform, they provide an accurate representation of the real data set. Low error in the frame time prediction and high values of F-statistic means the model is useful. The same templates can be used for execution time prediction of CPU-bound application. More examples about this will be added in the next revision. To demonstrate the power models, the tool has functions that predicts the power consumption of the CPU and GPU given their frequencies, activity ratio, dynamic capacitance, number of active cores and temperature.