[Solved]CUDA Memory Statistics

Image below is the memory statistics that i found in the website:
http://docs.nvidia.com/gameworks/index.html#developertools/desktop/nsight/analysis/report/cudaexperiments/kernellevel/memorystatistics.htm%3FTocPath%3DDeveloper%2520Tools%7CDesktop%2520Developer%2520Tools%7CNVIDIA%2520Nsight%2520Visual%2520Studio%2520Edition%7CNVIDIA%2520Nsight%2520Visual%2520Studi

I have 2 questions here:

  1. What is the difference between System memory and device memory here?
    Could GPU kernel access Host memory directly?
  2. Where is statistics of Constant Memory here? Does it similar to Texture memory?

system memory is host memory. device memory is GPU memory.

Yes, a GPU kernel can access host memory “directly” (over PCIE). The methodology is called zero-copy and there are sample codes demonstrating the concepts.

Constant memory is basically a cache, that is backed by a section of device memory. The path would be similar to Texture or Global RO, but I don’t think there are any metrics to track constant memory access.

http://docs.nvidia.com/cuda/profiler-users-guide/index.html#metrics-reference

This question:

http://stackoverflow.com/questions/37732735/nvprof-option-for-bandwidth

shows how some metrics map to the above diagram.

Thanks, txbob,.
Your answer is clear!