Nsight Visual Studio and Tesla P100 PCIE

Hey,

We’ve recently received our first Tesla P100 16GB PCIE cards. However they do not seem to be supported by the latest Nsight (5.2) as confirmed by this page https://developer.nvidia.com/nsight-visual-studio-edition-supported-gpus-full-list.

When can we expect proper support for the Tesla P100 in Nsight? Are there any alternative profiler that would work in the meantime?

Cheers,
Tanguy

Hi, GPSnoopy

Tesla P100 not supported by Nsight 5.2.

Of course, there is an alternative profiler called Visual Profiler in cuda toolkit.
download: CUDA Toolkit 11.7 Update 1 Downloads | NVIDIA Developer
help-doc: http://docs.nvidia.com/cuda/profiler-users-guide/index.html#axzz4WAmw90xg

We’ve tried several programs on the P100 and they work fine on their own. I had a go with the Visual Profiler today and it seems to work, which is good news.

When we try to run them in Nsight VSE, they run fine, all indicators are green, but Nsight then says that no kernel launches were detected. It’s like it’s captured no events.

We’re using latest Tesla P100 drivers for 2012R2 (376.33) with the latest Nsight 5.2 from November. The code runs as a non-admin, using Alea 3.0.2 in C#.

This is the first computer that we’ve set up with the P100. I’ll go over the installation to make sure the drivers, SDK and Nsight are installed correctly. Thinking of it, I’ll also try a C++ example to make sure it’s not Alea specific.

Will keep you posted on my progress.

Hi GPSnoopy,

Nsight 5.2.0.16321 (November release) does not support GP100 hardware, including the Tesla P100.

Please continue to use the CUDA Visual Profiler from the CUDA toolkit as veraj mentioned.

Sorry for the confusion.

Hi,

Thanks for the clarification!

Do you guys have a rough estimate on when Nsight will support the P100?

Cheers,
Tanguy

GPSnoopy

Currently, there is no plan to support GP100 profiling, including the Telsa P100, in NsightVSE.
Debugging of GP100, including Tesla P100, will be supported in an upcoming release.
Please use the CUDA Visual Profiler from the CUDA toolkit for profiling.

https://developer.nvidia.com/cuda-profiling-tools-interface
https://devblogs.nvidia.com/parallelforall/cuda-pro-tip-nvprof-your-handy-universal-gpu-profiler/

Hi rbischof,

Thanks for the complete answer.

Nsight allows to profile but also debug on Windows. This is a useful feature. I also find Nsight easier and quicker to use than the standalone visual profiler.

I’m confused how NVIDIA took the decision of supporting a $99 GeForce in Nsight but thought that nobody would want to use it when they splash $7000+ on a Tesla P100 (keep in mind the Tesla K40 is supported).

If there is ever an internal ticket at NVIDIA for P100 support in Nsight. Please add a +1 vote on my behalf.

Cheers,
Tanguy

Hi rbischof,

Is there a specific technical reason why GP100 profiling is not planned for NSightVSE? We’ve upgraded from Quadro K5200s to GP100s in our research lab. We used to profile the GPUs on NSightVSE counting the #of FP32/FP64 operations and memory transaction loads pretty easily with K5200s (even developed our own interface tools based on NSightVSE :/). Assuming you won’t change your mind about supporting NSightVSE for GP100 (hopefully you will though!), is it possible that we can somehow achieve these two features without NSightVSE on Win10 x64?

Thanks,
Burak