GPU Memory usages after cusolverMgDeviceSelect

Accelerated Computing GPU-Accelerated Libraries
1

I still try to study about cusolverMG library.

My workstation has 4 K80C cards. nvidia-smi command tells me that 11441MB memory is available for each GPU core. I’ve tried the example code in cuSOLVER :: CUDA Toolkit Documentation. I’ve watched the memory usage while running the code and found that core 0 uses 907MB and the other 7 cores use 112MB memory after run cusolverMgDeviceSelect routine. I didn’t allocate any memory at the moment. My question is as follows:

  1. What does the cusolverMgDeviceSelect do internally?

  2. Why does it allocate memory inhomogeneously?

Please help.

2

Did you ever get an answer to this question? I observed a similar loss of memory when I ran cusolverMgDeviceSelect.

3

I didn’t get any answers. Further, my system does not run the example code in multiple GPU simultaneously. So I’ve given up to solve the problem.

4

OK, thanks for the info.

5

FYI, I submitted a bug through the development portal. It turns out that it was a fixable bug. I was just informed that it will be fixed in the “next release” - I assume this means CUDA 11.0+

6

Oh, thanks to notice it.