Assessment of General Purpose GPU Systems in Real-Time Control

Abstract

The recent advances in GPU technology is offering great prospects in computation. However, the penetration of the GPU technology in real-time control has been somewhat limited due to two main reasons: 1) control algorithms for real-time applications involving highly parallel computation are not very common in practical applications and 2) the excellent performance in computation of GPUs is paid for by a penalty in memory transfer. As a consequence, GPU applications for real-time controls suffer from an often unacceptable latency. We present the factors that affect the performance of GPUs in real-time applications in fusion research in order to provide some hints to designers facing the option of using either a multithreaded, multicore CPU application or a GPU. In particular, we consider GPU usage in two common use cases in real-time applications in fusion research: dense matrix-vector multiplication for large state space-based control and online image analysis for feature extraction in camera-based diagnostics. Two applications mimicking the two use cases have been developed using the Tesla K40 GPU architecture, and the performance results are reported.