GPGPU performance and power estimation using machine learning

Abstract

Graphics Processing Units (GPUs) have numerous configuration and design options, including core frequency, number of parallel compute units (CUs), and available memory bandwidth. At many stages of the design process, it is important to estimate how application performance and power are impacted by these options. This paper describes a GPU performance and power estimation model that uses machine learning techniques on measurements from real GPU hardware. The model is trained on a collection of applications that are run at numerous different hardware configurations. From the measured performance and power data, the model learns how applications scale as the GPU's configuration is changed. Hardware performance counter values are then gathered when running a new application on a single GPU configuration. These dynamic counter values are fed into a neural network that predicts which scaling curve from the training data best represents this kernel. This scaling curve is then used to estimate the performance and power of the new application at different GPU configurations.Over an 8× range of the number of CUs, a 3.3× range of core frequencies, and a 2.9× range of memory bandwidth, our model's performance and power estimates are accurate to within 15% and 10% of real hardware, respectively. This is comparable to the accuracy of cycle-level simulators. However, after an initial training phase, our model runs as fast as, or faster than the program running natively on real hardware.

DeLTA: GPU Performance Model for Deep Learning Applications with In-Depth Memory System Traffic Analysis

“…To analyze the performance bottlenecks of GPU-accelerated data parallel workloads, many GPU performance models have been proposed [8]- [11], [24]- [26]. In particular, prior models do in-depth analysis of the parallel GEMM and show high accuracy [8], [9].…”

Energy efficient execution of heterogeneous applications

“…Sethia et al [26] built a dynamic power management scheme and simulated them using GPGPU-Sim [27]. Wu et al [28] built GPU performance and power prediction models using Artificial Neural Networks. They found a set of hardware counters that could be used for such prediction schemes.…”

“…The performance counters are collected at an intermediate frequency value, since measuring counters at extreme frequencies results in collecting a less-representative behavior, as the application behavior can change drastically when operating at extreme frequencies [28].…”

“…Our performance counter selection is based on our own analysis, as well as strategies presented in previous studies [28,23]. …”

“…Such a scheme has been discussed previously in [28], but has not been implemented or evaluated on real hardware. In addition, we could study more fine-grained DVFS control using simulators such as Multi2Sim [36] or gem5-gpu [37].…”