Up-to-date instructions for enabling GPU acceleration

Finally, after three previous failed attempts, was able to install and experience the GPU acceleration benefits with my new/old XPS8940 PC with Nvidia RTX3060 12GB GPU with 64GB of RAM.

P.S. had to tell my wife a small fib to buy this open box PC to get the Nvidia graphics card and the added CPU processing power, and now will have more time to spend with her since the processing times will be significantly reduced - haha...

What I learned:
  • Don't update your Game Ready Driver - I was proud to be on the current version (536.67 July 18, 2023 release date) and had to down-date the driver to 472.12 (for the RTX3060) and uninstalled new driver. Couldn't find the older driver but used the one that was loaded in the operating system, and so I went with it. Not sure if this was the root cause, but it worked. Will need to make sure this driver doesn't get automatically updated in the future.
  • If you've previously installed (and installation failed) the GPU acceleration application, tool kit, code and variables, carefully and thoroughly remove all remnants of past installations. Not sure which of the remnants created issues with each attempt to re-install the tools but found that removing everything seemed to be one of the critical success factors.

From others, will be interested to hear if subsequent driver updates (current version for my GPU is 536.67) will work with the current PixInsight GPU configuration.

Remove (inspect) all of the previous environmental variables (some remained from other GPU acceleration installation instructions that failed to work for me).

Followed the RT-- step-by-step installation process carefully - easy to make typographical errors (fat fingered the environment variables twice).

The performance improvement was absolutely eye popping...

NXT on the Crescent Nebula (281,251KB file): Before processing time - 9:04 minutes; After processing time - 11.04 seconds

BXT processing time after acceleration on the same file: 35.08 seconds

SXT: 27.52 seconds

BXT on Crescent Nebula - 2x Drizzle (1,125,328 KB file): 1:41 minutes versus 23+ minutes without GPU acceleration!

Riku - if you read this post, let me know how I can buy you a coffee or some other beverage of your choice - your GPU acceleration guide was absolutely the best!
 
Finally, after three previous failed attempts, was able to install and experience the GPU acceleration benefits with my new/old XPS8940 PC with Nvidia RTX3060 12GB GPU with 64GB of RAM.

P.S. had to tell my wife a small fib to buy this open box PC to get the Nvidia graphics card and the added CPU processing power, and now will have more time to spend with her since the processing times will be significantly reduced - haha...

What I learned:
  • Don't update your Game Ready Driver - I was proud to be on the current version (536.67 July 18, 2023 release date) and had to down-date the driver to 472.12 (for the RTX3060) and uninstalled new driver. Couldn't find the older driver but used the one that was loaded in the operating system, and so I went with it. Not sure if this was the root cause, but it worked. Will need to make sure this driver doesn't get automatically updated in the future.
  • If you've previously installed (and installation failed) the GPU acceleration application, tool kit, code and variables, carefully and thoroughly remove all remnants of past installations. Not sure which of the remnants created issues with each attempt to re-install the tools but found that removing everything seemed to be one of the critical success factors.

From others, will be interested to hear if subsequent driver updates (current version for my GPU is 536.67) will work with the current PixInsight GPU configuration.

Remove (inspect) all of the previous environmental variables (some remained from other GPU acceleration installation instructions that failed to work for me).

Followed the RT-- step-by-step installation process carefully - easy to make typographical errors (fat fingered the environment variables twice).

The performance improvement was absolutely eye popping...

NXT on the Crescent Nebula (281,251KB file): Before processing time - 9:04 minutes; After processing time - 11.04 seconds

BXT processing time after acceleration on the same file: 35.08 seconds

SXT: 27.52 seconds

BXT on Crescent Nebula - 2x Drizzle (1,125,328 KB file): 1:41 minutes versus 23+ minutes without GPU acceleration!

Riku - if you read this post, let me know how I can buy you a coffee or some other beverage of your choice - your GPU acceleration guide was absolutely the best!

Good to hear you got it working!
The cuDNN driver requirements have not changed for the latest version (8.9.3 still has the same 452.39 minimum), so I don't expect there to be any incompatibilities. I keep my drivers up-to-date, and haven't had any issues.

There's a PayPal donate link on the bottom of the tutorial (not visible in mobile view), if you want to support. 🙂
 
Finally, after three previous failed attempts, was able to install and experience the GPU acceleration benefits with my new/old XPS8940 PC with Nvidia RTX3060 12GB GPU with 64GB of RAM.

P.S. had to tell my wife a small fib to buy this open box PC to get the Nvidia graphics card and the added CPU processing power, and now will have more time to spend with her since the processing times will be significantly reduced - haha...

What I learned:
  • Don't update your Game Ready Driver - I was proud to be on the current version (536.67 July 18, 2023 release date) and had to down-date the driver to 472.12 (for the RTX3060) and uninstalled new driver. Couldn't find the older driver but used the one that was loaded in the operating system, and so I went with it. Not sure if this was the root cause, but it worked. Will need to make sure this driver doesn't get automatically updated in the future.
  • If you've previously installed (and installation failed) the GPU acceleration application, tool kit, code and variables, carefully and thoroughly remove all remnants of past installations. Not sure which of the remnants created issues with each attempt to re-install the tools but found that removing everything seemed to be one of the critical success factors.

From others, will be interested to hear if subsequent driver updates (current version for my GPU is 536.67) will work with the current PixInsight GPU configuration.

Remove (inspect) all of the previous environmental variables (some remained from other GPU acceleration installation instructions that failed to work for me).

Followed the RT-- step-by-step installation process carefully - easy to make typographical errors (fat fingered the environment variables twice).

The performance improvement was absolutely eye popping...

NXT on the Crescent Nebula (281,251KB file): Before processing time - 9:04 minutes; After processing time - 11.04 seconds

BXT processing time after acceleration on the same file: 35.08 seconds

SXT: 27.52 seconds

BXT on Crescent Nebula - 2x Drizzle (1,125,328 KB file): 1:41 minutes versus 23+ minutes without GPU acceleration!

Riku - if you read this post, let me know how I can buy you a coffee or some other beverage of your choice - your GPU acceleration guide was absolutely the best!
I just finished putting together a new computer largely optimized for image processing (PI and PS). I followed the CUDA setup instructions on the RC-ASTRO site, and everything went perfectly well. Thought it was working at first, but I was fooled because the new machine is inherently much faster than the old one. The CPU was still doing all the work (despite the non-CPU tensorflow library being used by PI). The only procedural difference between Russ's instructions and the ones here involve limiting the CUDA Toolkit installation to just the libraries. Can't see why that would matter, but I uninstalled the toolkit and reinstalled with just the libraries (the uninstall removed a lot of hooks to my Visual Studio installation), and now it works fine, blazingly fast with the RC tools and StarNet. No idea why one worked and the other didn't, but an additional data point for anybody having problems.

I use the NVidia Studio driver installation, not the Game one. No idea what the actual difference is. I have the latest one, however, and the CUDA installation checks for compatibility (it claims it will replace the driver if necessary) and found no issues. So my system is working fine with the new driver and the CUDA 11.8 install.
 
Happy to report another successful installation with remarkable results, RTX 2070 GPU, BXT around 8 minutes down to 35 seconds - with no extra hardware expenditure! 🥳
Thank you so much for providing easy installation instructions!
 
I have a new RTX 4070 Ti and was struggling to get CUDA to work in PixInsight. One last helper if anyone is struggling with later Nvidia GPU's. There was one additional thing I had to do to get CUDA working with PixInsight, and that is to disable Hardware-accelerated GPU scheduling.

Go to Settings -> System -> Display -> Graphics settings (down the bottom of page) -> Hardware-accelerated GPU scheduling -> Switch to "Off".

Reboot PC

Once I did this I noticed CUDA showing back up in the performance graph drop down for the NVIDIA RTX 4070 Ti I have, and I see PixInsight using CUDA for starnet,etc..
 
I have a new RTX 4070 Ti and was struggling to get CUDA to work in PixInsight. One last helper if anyone is struggling with later Nvidia GPU's. There was one additional thing I had to do to get CUDA working with PixInsight, and that is to disable Hardware-accelerated GPU scheduling.

Go to Settings -> System -> Display -> Graphics settings (down the bottom of page) -> Hardware-accelerated GPU scheduling -> Switch to "Off".

Reboot PC

Once I did this I noticed CUDA showing back up in the performance graph drop down for the NVIDIA RTX 4070 Ti I have, and I see PixInsight using CUDA for starnet,etc..

I thought that that switch only affected the visibility of the graph, not the actual performance.

Edit:
A quick test:
Hardware-accelerated GPU scheduling disabled: StarXTerminator runs in 16s, things show up in the Cuda graph
Hardware-accelerated GPU scheduling enabled: StarXTerminator runs in 16s, things show up in the 3D graph (no Cuda graph available)
 
Last edited:
Back
Top