It doesn’t appear that the version of CUDA installed on the xavier (CUDA 10) is actually support by Nsight systems? Is that correct? Is there no way to get both CPU and GPU profiling informations on this board?
Hi,
Have you indicated the Xavier GPU architecture and CUDA 10.0?
[url]https://devtalk.nvidia.com/default/topic/1010551/jetson-tx2/using-nsight-eclipse-edition-with-opencv-for-cross-compilation-/post/5175091/#5175091[/url]
Please use “-gencode arch=compute_72,code=sm_72” for Volta.
Thanks.
Hmm, that only affect’s compilation I think. I believe I have my compilation setup correctly and everything runs correctly and can be profiled with nvprof. However, when I try to use the NSight systems to capture CUDNN/CUDA calls, it doesn’t detect any. I noticed that the injection lib files it hasn’t doesn’t support CUDA 10 or the CUDNN version installed and NSight Compute’s injection libs also do not support CUDA 10(or arm processor targets either I think).
Hi,
We haven’t received other issues about the regression of Nsight on CUDA 10.
We will check this internally.
Will update information with you later.
Thanks.
Hi,
Um think this is the same unresolved issue.
https://devtalk.nvidia.com/default/topic/1044768/nsight-system-not-generating-cuda-traces-for-agx-xavier/
I don’t think Nsight Systems of Nsight Compute supports any CUDA 10 on ARM.
Thanks,
Thanks for your information.
We will check this with internal team and update information with you.
