GPU rendering core load percentage

Accelerated Computing CUDA CUDA Programming and Performance
1

Dear all,

I am developing a GPU CUDA code and running that on TITAN V. When I run the code and monitor the GPU load by the software GPU-Z, it is around 50%. How can I increase the GPU load or use 100% of the GPU card?

Thank you for your help in advance.

2

Profile your code, and modify your code in such a way that the code timeline is entirely occupied by kernel calls. To a first order approximation, the GPU-Z utilization number is reporting a percentage of the timeline during which one or more GPU kernels are active.

3

Thank you for the quick reply.
So, playing with the value of threads and blocks don’t affect the GPU load? I need to modify or optimize the code?
I think the CUDA samples such as fluidsGL have been optimized, but even for this sample, the GPU load is 50%.

4

CUDA samples generally haven’t been optimized for performance. Their primary purpose is to demonstrate a particular technique or method.

A short sample code that does just one thing, often perhaps launching a single kernel, may not score very well on this measurement.

Threads and blocks don’t affect this particular GPU load measurement. The fluidsGL code is a CUDA/OpenGL interop code, which means it is alternating between launching a kernel and handing off the computed data to OpenGL, then launching a kernel again, then handing off the data again. I think you might learn a lot by learning to use a profiler.

If you want to see that number close to 100%, I would use the method I already described. Launch even a trivial kernel that occupies most of the duration of the program.

Try compiling and running this code, and see what GPU Z reports:

__global__ void k(){
        long long start = clock64();
        while (clock64() < start+10000000000ULL);
}

int main(){

        k<<<1,1>>>();
        cudaDeviceSynchronize();
}

It should run for about 5-10 seconds, and I expect GPU-Z to report 100% utilization for much of that time period.

5

Thank you so much.