In the CUDA visual profiler, version 5, I know that the "gld/st_requested_throughput" are the requested memory throughput of application. However, when I try to find the actual throughput of hardware, I am confused because there are two pairs of metrics which seem to be qualified, and they are "gld/st_throughput" and "dram_read/write_throughput". Which pair is actually the hardware throughput? And what does the other serve as?
gld/st_throughput includes transactions served by the L1 and L2 caches. While dram_read/write_throughput is the throughput between L2 and device memory. So, each global memory access counts towards gld/st_throughput but only requests that missed both the L1 and L2 caches count towards dram_read/write_throughput.
I haven't found a good overview of the counters anywhere. Wish NVIDIA would provide that...
来源：https://stackoverflow.com/questions/13840328/whats-the-difference-between-gld-st-throughput-and-dram-read-write-throughpu