The description in the user guide is too short to make sense to me.
I don’t have experience with CUDA programming.
If you don’t have a lot of experience with CUDA, there is no reason for you to be using the --cudabacktrace=other.
If you are just looking to find out how your CUDA calls are happening on the CPU and GPU, you only need to call “nsys profile app.name”. The default parameters include tracing NVTX (NVIDIA’s annotation library), OSRT (operating system run time information), CUDA, and OpenGL.
This will give you information that will show you the CPU side CUDA API calls as well as show you when those CUDA kernels were running on the GPU side.
–cudabacktrace should only be run at all if you need more clarity about where in your code those calls are being made from, and for the most part, if you need it, you should be using the “all” category.