Hello…
When i profile my arithmetic intensive kernel its IPC is 1.99 on my tesla card(2075). Profiler says Max IPC is 2.0 . Does this mean the GPU cannot do anymore computation during the kernel execution time , because IPC is almost the peak
Also the only way to optimize now is reduce the total arithmetic instructions?
Also on an another card with compute capability 2.1 the IPC shown is 2.4, where max IPC is 4.0. Does this mean there is still room for more arithmetic operation in this card but not the other?
Regards,
Vinod
I think It almost achieve the peak. Devices with compute capability 2.1 has two scheduler, and each scheduler could issue two instructions per cycle. However, it take two cycles for a warp to complete , as each instruction only reach half of the warp (16 cores). Thus even schedulers work at its maximum, it will limited by the number of cores, as there are 3 groups of half core. i.e IPC 3 is the maximum