Similar behaviour here. But if you click on GPU to view the details you can see some activity in the memory copy section. My guess is that the algorithm itself uses a lot of memory accesses but not so much computations. That's generally not ideal for GPU acceleration. However, the application seems to be some kind of straightforward build from the python code so the GPU code most likely comes from python libraries. So there definitely is a room for optimization but it would probably require rewriting the app from scratch in CUDA C/C++.