Back to forums

Enabling GPU usage on an NVIDIA card for RC-Astro tools

Help Needed Troubleshooting
12 replies•392 views
Gordon Pegue avatar
Gordon Pegue avatar

Trying to get this to work has been entirely hit or miss.

I’ve tried simply the “experimental” repository method outlined on the RC-astro pages.

That worked for a while but now my RC-Astro plugins BXT, NXT and STX are just slow, taking minutes at a time.

So I thought, what the heck, I’ll try the manual method outlined here: Link

Followed it exactly but I’m seeing no difference in the Windows 11 Task Manager GPU performance for my RTX4060 card.

Is there a definitive solution to my dilema?

What method is everyone else (that uses PI installed on a windows 11 device) doing???

Thanks,

CS

bigCatAstro avatar
bigCatAstro avatar

*deleted

Brian Puhl avatar
Brian Puhl avatar

You’re not using a laptop are you? Nvidia Optimus could be forcing integrated graphics instead of the GPU.

When I first installed the tensorflow tools, I followed that same rikutalvio guide years ago. I suspect you’re missing something, maybe just overlooked. Try a restart if you haven’t.

bigCatAstro avatar
bigCatAstro avatar

Gordon Pegue · Sep 7, 2025 at 03:16 AM

Trying to get this to work has been entirely hit or miss.

I’ve tried simply the “experimental” repository method outlined on the RC-astro pages.

That worked for a while but now my RC-Astro plugins BXT, NXT and STX are just slow, taking minutes at a time.

So I thought, what the heck, I’ll try the manual method outlined here: Link

Followed it exactly but I’m seeing no difference in the Windows 11 Task Manager GPU performance for my RTX4060 card.

Is there a definitive solution to my dilema?

What method is everyone else (that uses PI installed on a windows 11 device) doing???

Thanks,

CS

What version of CUDA are you currently running? I’m running 13.0 and it took several hoops to jump through to get it fully calibrated to run with Pixinsight. I was getting tensorflow errors though (so not the same as you), which prompted me to upgrade.

BXT is clocking for me at sub 30 seconds with 13.0.

Gordon Pegue avatar
Gordon Pegue avatar

@Brian Puhl :

No, not a laptop. A solid dell Precision fixed workstation that is otherwise quite zippy.

I rebooted after completing the steps in that guide. I’m retired IT so it was easy as pie.

@bigCatAstro :

Per the guide, I’ve got CUDA Toolkit 11.8.0 installed…

13.0 was offered when I logged into my NVIDIA developer account.

If I scrub everything and install Toolkit 13.0, what are the versions of the other downloads I’d need?

Gotta cheat sheet for the requirements??

Best,

Gordon

D. Jung avatar
D. Jung avatar

Did you follow this

https://pixinsight.com/forum/index.php?threads/experimental-tensorflow-gpu-acceleration-repository.22325/

You actually don't need to install anything really. At least that's how it worked for me with my 4060.

bigCatAstro avatar
bigCatAstro avatar

Gordon Pegue · Sep 7, 2025 at 06:04 AM

@Brian Puhl :

No, not a laptop. A solid dell Precision fixed workstation that is otherwise quite zippy.

I rebooted after completing the steps in that guide. I’m retired IT so it was easy as pie.

@bigCatAstro :

Per the guide, I’ve got CUDA Toolkit 11.8.0 installed…

13.0 was offered when I logged into my NVIDIA developer account.

If I scrub everything and install Toolkit 13.0, what are the versions of the other downloads I’d need?

Gotta cheat sheet for the requirements??

Best,

Gordon

I followed this guide since I was still unable to get the GPU to function with other guides: https://www.rc-astro.com/gpu-acceleration-for-ai-powered-tools/

I was missing the cuDNN files and zlib specifically. Some of the steps are now out dated, e.g. the zlib DLL file name and the libnvvp folder for earlier versions of CUDA has been rolled into 13.0, the path variable step for that folder is not needed.

Mind you, the issue I was experiencing is probably different since you actually had a working system.

Gordon Pegue avatar
Gordon Pegue avatar

D. Jung · Sep 7, 2025, 07:12 AM

Did you follow this

https://pixinsight.com/forum/index.php?threads/experimental-tensorflow-gpu-acceleration-repository.22325/

You actually don't need to install anything really. At least that's how it worked for me with my 4060.

That’s actually what I did first @D. Jung and it worked fine for a while… (a couple of months)

Gordon Pegue avatar
Gordon Pegue avatar

bigCatAstro · Sep 7, 2025, 01:48 PM

I followed this guide since I was still unable to get the GPU to function with other guides: https://www.rc-astro.com/gpu-acceleration-for-ai-powered-tools/

I was missing the cuDNN files and zlib specifically. Some of the steps are now out dated, e.g. the zlib DLL file name and the libnvvp folder for earlier versions of CUDA have been rolled into 13.0, the path variable step for that folder is not needed.

Mind you, the issue I was experiencing is probably different since you actually had a working system.

That guide echoes the rikutalvio sequence - without identifying specific versions of the cuDNN, zlib & tensorflow components.

Is zlib now rolled up in CUDA toolkit 13.0?

If so, my next steps are to remove CUDA 11.8, scrub the file system of the older cuDNN and zlib files, scrub the environment variables and try CUDA Toolkit 13.0, cuDNN 9.13.0 and tensorflow 2.10.

Will advise,

CS

bigCatAstro avatar
bigCatAstro avatar

Gordon Pegue · Sep 7, 2025 at 05:30 PM

bigCatAstro · Sep 7, 2025, 01:48 PM

I followed this guide since I was still unable to get the GPU to function with other guides: https://www.rc-astro.com/gpu-acceleration-for-ai-powered-tools/

I was missing the cuDNN files and zlib specifically. Some of the steps are now out dated, e.g. the zlib DLL file name and the libnvvp folder for earlier versions of CUDA have been rolled into 13.0, the path variable step for that folder is not needed.

Mind you, the issue I was experiencing is probably different since you actually had a working system.

That guide echoes the rikutalvio sequence - without identifying specific versions of the cuDNN, zlib & tensorflow components.

Is zlib now rolled up in CUDA toolkit 13.0?

If so, my next steps are to remove CUDA 11.8, scrub the file system of the older cuDNN and zlib files, scrub the environment variables and try CUDA Toolkit 13.0, cuDNN 9.13.0 and tensorflow 2.10.

Will advise,

CS

Just finally got to edit my post, the libnvvp folder contents have been rolled into cuda 13.0. The zlib dll file still needs to be downloaded from zlib. The file isn’t called zlibapi (or whatever it was called) anymore.

Gordon Pegue avatar
Gordon Pegue avatar

Welp…

After scrubbing the previous CUDA stuff, I installed these:

cuda_13.0.1_windows.exe

cudnn_9.13.0_windows.exe

Copied the tensorflow dll from libtensorflow-gpu-windows-x86_64.2.10.0.zip into the PixInsight bin folder.

I did nothing about the zlib file…. Environment variables point to correct locations.

Now PI crashes outright when I try and launch any RC-Astro plugin:

For example, I load a masterlight XISF and run BXT. The PI console says Initializing… and then BOOM gone.

This sucks… Oh, to have a mac.

Why do I hafta reinvent the wheel here? Somebody has to have a working setup…

Gordon Pegue avatar
Gordon Pegue avatar

YAY! finally solved the puzzle!!

First, knowing that I was trying to use tensorflow 2.10, I found this page which has a matrix of what the other component versions needed to work with 2.10: https://www.tensorflow.org/install/source#gpu

So I downloaded (per the matrix) these components:

cuda_11.2.0_460.89_win10.exe;

cudnn-10.2-windows10-x64-v8.1.0.77 dot zip (dunno why this GUI wants to turn that file name into a link).

Already had the zlib123dllx64 dot zip containing the zlibwapi.dll and the libtensorflow-gpu-windows-x86_64-2.10.0 dot zip containing the tensorflow.dll file.

Did an express install of the CUDA toolkit, installing just the libraries and followed the rikutalvio steps to finish the manual setup. Rebooted the PC.

Launched PI.

Held my breath as I launched BXT on a 112MB red masterlight of NGC1365…

Boy did BXT haul it! What was taking minutes previously, took a little more than 20 seconds!!

So in summary, I hope this response helps some other poor soul who is struggling to make this work on a windows 11 box.

Read noise Astrophotography avatar
Read noise Astrophotography avatar

Gordon, I ran into the same headaches it’s all about CUDA versions. RC-Astro tools (BXT, NXT, STX) still want CUDA 11.8, but many of us (me included) need CUDA 12 for other workloads like LLMs.

PixInsight doesn’t “know” how to juggle both, so if CUDA 12 is first in your PATH, the plugins crawl or fail. The fix is to wrap PixInsight in a custom launcher so it only sees CUDA 11.8.

Here’s the batch file I use on my Windows 11 rig (RTX 5080 here, but works fine on 4060 too). Save this as PixInsight_CUDA11.bat, put it on your desktop, and use it to start PI:

Drop these in to your PIX bin folder

cublas64_11.dll

  • cufft64_10.dll

  • cusolver64_11.dll

  • cudnn64_8.dll




    @echo off

setlocal

:: Path to PixInsight bin folder

set "PI_BIN=C:\Program Files\PixInsight\bin"

:: Path to CUDA 11.8 binaries

set "CUDA11_BIN=%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin"

:: Make sure PI’s local DLLs come first

set "PATH=%PI_BIN%;%PATH%"

:: Strip out CUDA 12.* from PATH for this process

for %%V in (0 1 2 3 4 5 6 7 8 9) do (

set "PATH=%PATH:%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.%%V\bin;=%"

for %%M in (0 1 2 3 4 5 6 7 8 9) do (

set "PATH=%PATH:%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.%%V%%M\bin;=%"

)

)

:: Add CUDA 11.8 bin (fallback for any missing DLLs)

if exist "%CUDA11_BIN%\cudart64_110.dll" set "PATH=%PI_BIN%;%CUDA11_BIN%;%PATH%"

:: TensorFlow behaviour (optional but helps stability)

set "TF_CPP_MIN_LOG_LEVEL=0"

set "TF_FORCE_GPU_ALLOW_GROWTH=TRUE"

set "CUDA_VISIBLE_DEVICES=0"

set "TF_LOG_DEVICE_PLACEMENT=1"

cd /d "%PI_BIN%"

start "" "%PI_BIN%\PixInsight.exe"

endlocal

Helpful