Performance and power evaluation of SH-X3 multicore system

We have developed an embedded processor that supports asymmetric multiple processor (AQMP), symmetric multiple processor (SMP), and an AMP/SMP hybrid system. It contains four SH-X3 cores used to support cache coherency from that obtained using an SH-X2 core. In this paper, we evaluate the following three techniques to improve the processing performance and reduce the power consumption in parallel processing in the processor. The first technique is snoop controller (SiNC) to improve cache coherency performance. The performance overhead by snoop is decreased up to 0.1% when SPLASH-2 is executed. The second technique is detection and resolution of synonym problems so that we may not use the page coloring for page table management. The processes handling time in Linux is reduced by 29.4% compared with the case solved the problem with software. The third technique is the individual core clock frequency and the light sleep mode which is used to maintain the cache coherency even when the cores are stopped, to reduce the power consumption. The energy is decreased by 5.2% and 4.5%, respectively. As a result, the SH-X3 core achieved a performance that has scalability proportional to 0.72-0.93 times the number of cores and a power saving of 4.5-44.0% without increasing the execution time.