Fig. 1

A simplified diagram of a GPU used in this study. A GPU has 30 multiprocessors, and each multiprocessor has eight processor cores (240 cores in total). The shared memory can be accessed from every core in a multiprocessor, while registers are local to each core. Several tens of registers are typically available for a thread. Latency of the shared memory is two orders of magnitude lower than that of global memory (Abdelkhalek et al., 2009; Micikevicius, 2009).