My question is on how should one interpret the memory percentage of the utilization menu by nvidia-smi, to point them on the direction of improving their GPU code performance. My hardware is a Quadro 4000 and I am using CUDA 5.0. For most examples that I test with my code I get on average:
Gpu : 80 %
Memory : 20 %
However, for some specific example examples I have
Gpu : 39 %
Memory : 3 %
and consequently the performance drops significantly. However, the algorithm used to model both cases is the same.
Given the fact that the memory percentage is related to global memory, would you say that this number indicates a coalescing issue? Moreover, is there any way Nsight can help me with this?