Hello all,
I'll acknowledge up front I'm no expert with the CUDA stack. I had quite a bit of trouble setting up my windows 10 laptop with NVIDIA GPU for BXT and SXT. I couldn't get the latest version of CUDA Toolkit (v12.0) to work under multiple attempted configurations/versions of the other dependencies. After quite a bit of struggling, I found some additional helpful info, and now the GPU is functioning correctly. I post here in case it helps folks who are having similar problems.
In the end, what I found most useful was a web article that helped me understand the relationship of version dependencies between GPU "compute capability" (i.e. number) of my specific NVIDIA GPU, the CUDA Toolkit version (including graphics driver), and the corresponding CuDNN and TensorFlow version numbers. While some may get lucky and find the latest versions work, others may need to be more careful with the version dependencies. I went "back to basics" and aligned these versions to my hardware per the info in the post below. Not all the info is needed for PixInsight use; just the dependencies listed above and then application as per the instructions given in PixInsight BXT documentation. Part 1 and Part 2 of the following article may help.
https://medium.com/analytics-vidhya/step-by-step-guide-to-setup-gpu-with-tensorflow-on-windows-laptop-c84634f59857
Note that if you've already installed the wrong (incompatible) versions of these items, you'll have environment variables to remove or change (set by CUDA Toolkit installation) in addition to the unwind of components. I hope this post helps others to avoid the pain/frustration that I experienced trying to get the GPU working. The speed increase is worth the trouble once the dependencies are discovered and sorted.
Cheers, and CS, Doug S.
I'll acknowledge up front I'm no expert with the CUDA stack. I had quite a bit of trouble setting up my windows 10 laptop with NVIDIA GPU for BXT and SXT. I couldn't get the latest version of CUDA Toolkit (v12.0) to work under multiple attempted configurations/versions of the other dependencies. After quite a bit of struggling, I found some additional helpful info, and now the GPU is functioning correctly. I post here in case it helps folks who are having similar problems.
In the end, what I found most useful was a web article that helped me understand the relationship of version dependencies between GPU "compute capability" (i.e. number) of my specific NVIDIA GPU, the CUDA Toolkit version (including graphics driver), and the corresponding CuDNN and TensorFlow version numbers. While some may get lucky and find the latest versions work, others may need to be more careful with the version dependencies. I went "back to basics" and aligned these versions to my hardware per the info in the post below. Not all the info is needed for PixInsight use; just the dependencies listed above and then application as per the instructions given in PixInsight BXT documentation. Part 1 and Part 2 of the following article may help.
https://medium.com/analytics-vidhya/step-by-step-guide-to-setup-gpu-with-tensorflow-on-windows-laptop-c84634f59857
Note that if you've already installed the wrong (incompatible) versions of these items, you'll have environment variables to remove or change (set by CUDA Toolkit installation) in addition to the unwind of components. I hope this post helps others to avoid the pain/frustration that I experienced trying to get the GPU working. The speed increase is worth the trouble once the dependencies are discovered and sorted.
Cheers, and CS, Doug S.
