Efficient Data Encoding for Improving Fault Simulation Performance on GPUs
Jason G. Tong, Marc Boulé, Željko Žilić · 2013
Graphics Processing Units (GPUs) have recently gained widespread usage as an advanced parallel platform for accelerating compute intensive applications. One of the key factors for achieving maximal performance of GPU kernels is to ensure that the data used for computation is mapped independently. This implies that data duplication is needed, however, an efficient mapping of data is imperative as the on-board memory capacities on GPUs are limited. In this paper, we present a novel data encoding technique for creating an efficient data mapping for the application of fault simulation algorithms. Fault simulation requires a certain level of data dependency which creates a challenge for efficiently mapping the circuit data on the GPU's device memory. Based on our results, our memory optimization techniques were able to reduce the memory usage by 80% with speed ups reaching over 60× in the circuit benchmarks tested.