PixInsight GPU acceleration

[Peter Ward]

There have been posts elsewhere in this forum in relation to GPU acceleration of image processing using PixInsight and RC-Astro's AI based processing tools (Noise, Blur and Star removal).

These tools work remarkably well but can take several minutes to render
per image.

Nvidia's powerful Blackwell series GPU processor has recently been introduced but many self appointed "experts" on Facebook etc.
report they will not work with the RC-Asto suite.

This is simply not true.

However you will need to use certified Nvidia or Gigabyte GPU's. You can check to see what series/make of GPU is compatible with CUDA software on Nvidia's website.

I am using a Gigabyte 5070X

The easiest way to implement GPU acceleration of RC-Astro's tools is to add a new single line to the PixInsight repository (see RC-Astro's website for the details)

RC-Astro will then update Pixinsight with ALL of the tedious steps otherwise required. (e.g.Download Nvidia's CUDA, extra .dll's, enviroment variables, new Tensorflow.dll etc.)

I would however strongly recommend....before you do the above...rename the Tensorflow.dll file in your Pixinsight Bin folder to Tensorflow_CPU.dll (or similar) as a backup, in case all fails.

Also make a copy of the new Tensorflow.dll after the update above and save it as say TensorflowBAK.dll as future Pixinsight updates may overwrite the new GPU version.

When you first run any of the BlurX, StarX or NoiseX with a 5000 series card it will likely appear Pixinsight hangs at the image initialisation phase. Don't panic.
Let it think about it.
It might take ten minutes. (I had to wait six)

On subsequent processing, the time taken is nothing short of incredible.
A BluxX run that previously took 2 minutes 10 seconds to process was done in 4.2 seconds.

That's 30x faster than non-GPU processing times.

[joshman (Josh)]

RC's Pixinsight repositry for enabling CUDA acceleration has been a lifesaver.


I've been using the RC suite with GPU acceleration for a while now, and the speed boost is truly incredible. I usually setup a process container to run several of the processes in sequence, and the worst I can recall so far was ~10 minutes to run a sequence of BlurX, StarX, and Noise X on a single Drizzle 2x image from an IMX455 Mono FF Sensor. (~240MP image)


I dare not contemplate what that would've taken without GPU acceleration.

[g__day (Matthew)]

Solid advice Peter - having a single include file is so much easier than the old 12 step program requiring accounts with NVidia to download everything you require.

Backing up your tensorflow.dll file is very solid advice - I have done that before every major release of PI.

I mainly run my GPU based tasks - the Exterminators and Seti Astro's Cosmic Clarity Sharpen and Denoise on my RTX 4070 - it is about 10x - 15x times faster than my dual Xeons.

I have noticed when the Xterminators are first processing an image - they would take about 8 - 12 seconds to load it to the graphics card then the GPU based processing would complete in 5 - 15 seconds. Seti Astro's GPU enabled scripts take up to 15 - 20 seconds to load the image or the AI neural network - then take 10 - 20 seconds to process the image.

I just which the LN reference generation and local normalisation steps in WBPP were both GPU accelerated - that would be a real boon!