Up-to-date instructions for enabling GPU acceleration

I have a new computer that does not yet have all the required programs for AI work. Would following these same instructions apply for my situation?

thanks,
alex woronow
 
I don't have very deep understanding of what under the hood in computer graphics. I am using an older gaming computer with Geoforce GTX 1050. Would it be worthwhile to try and speed this up? I haven't seen any success stories for it yet.

thanks,
Al Force
 
RT-- said:
All instructions for Pixinsight CUDA acceleration I've seen are too old to cover the latest generation of GPUs, so I wrote a tutorial. This should work on anything from GTX900 to RTX4000-series.

Even if you already got it to work using an older version of CUDA, it's a worthwhile update that will give a hefty speed boost with some GPUs.

https://rikutalvio.blogspot.com/2023/02/pixinsight-cuda.html
Click to expand...
I had to update the driver for my GeForce GTX 1050 but when I tried to get the very latest drive it kept giving an error. So I downloaded and installed a less than current one (but still higher than 452.39 and that went OK. I followed the directions slowly and very carefully. When I tested in Pixinsight I saw no noticeable improvement and really didn't see anything in Task Manager. I did not see any graph area for CUDA or 3D but I never use 3D and it might not be set up. In the NVIDIA Control Panel I see CUDA and it show 640 cores. Any suggestions?

thanks,
Al
 
allen456 said:
I had to update the driver for my GeForce GTX 1050 but when I tried to get the very latest drive it kept giving an error. So I downloaded and installed a less than current one (but still higher than 452.39 and that went OK. I followed the directions slowly and very carefully. When I tested in Pixinsight I saw no noticeable improvement and really didn't see anything in Task Manager. I did not see any graph area for CUDA or 3D but I never use 3D and it might not be set up. In the NVIDIA Control Panel I see CUDA and it show 640 cores. Any suggestions?

thanks,
Al
Click to expand...

If you can't see GPU usage increasing, what does your CPU usage look like then?
 
For whatever reason, maybe rebooting, suddenly it is working. My NVIDIA GPU jumped way up and STX really flew. Looks like I'm good.
 
Tried it today and made the mistake of downloading the latest CUDA libraries - v12.2 rather than 11.8 - so no joy there; went back through the instructions and downloaded all right libraries and it is much faster now!

The only thing I noticed was it grabbing 11GB of RAM on the GPU (an old NVidia GTX 1080Ti) and it didn't release the memory once BlurXterminator was run - and held it until PI was closed. Is this expected behaviour? Not a particular problem but curious that it hold onto the GPU RAM...

Must say I love it when modern programs allow GPU loadsharing - so a huge thank you to the folks who made it happen!
 
i've seen tensorflow not release memory even on OSX, so i think this is a tensorflow issue. the calling process might have to terminate before tensorflow tears everything down. i even asked nikita about this (re: starnet, which is where i saw the behavior) and he was making all the right calls to tensorflow to destroy the graph and whatnot.
 
RT-- said:
All instructions for Pixinsight CUDA acceleration I've seen are too old to cover the latest generation of GPUs, so I wrote a tutorial. This should work on anything from GTX900 to RTX4000-series.

Even if you already got it to work using an older version of CUDA, it's a worthwhile update that will give a hefty speed boost with some GPUs.

https://rikutalvio.blogspot.com/2023/02/pixinsight-cuda.html
Click to expand...
I guess this could be a little dangerous to non techies? Otherwise I am not sure why this is not more widely known. The impact is increadible and a huge time saver. Thanks so much for taking the time to write this up for us.
 
I Attempted to enable GPU acceleration with CUDA for Pixinsight following the very successful writeup by Riku.

I followed along the procedure and verified everything more than once, and rebooted many times,

but it would not work ☹. So I bit the bullet and reinstalled everything finally resulting in success ! 😊



I have an NVIDIA RTX A2000 GPU with 12 GB RAM running driver 472.47 with 3328 CUDA cores

The AMD threadripper pro 5965wx 24 core, 48 thread CPU with compute capability of 8.6

128 GB RAM

Running 64 bit windows 11 , ver 22H2

For test comparison I ran starXterminator on a pic ;

On my old laptop with dual quad processors (8 threads) it took 11.5 min

On my new PC above without the mod (48 threads) it took 2.5 min

On my new PC above with the mod (GPU has 3328 cuda cores) it took 12 sec, the GPU never exceeded 10% load and CPU was in the noise barely active.

I am quite happy with the results ! Thank you Riku !! Not sure why it took a reinstall ?
Jim
 
RT-- said:
All instructions for Pixinsight CUDA acceleration I've seen are too old to cover the latest generation of GPUs, so I wrote a tutorial. This should work on anything from GTX900 to RTX4000-series.

Even if you already got it to work using an older version of CUDA, it's a worthwhile update that will give a hefty speed boost with some GPUs.

https://rikutalvio.blogspot.com/2023/02/pixinsight-cuda.html
Click to expand...
Finally! I've tried some other instructions and it didn't work, but this finally worked and the instructions were nice and clear. Thank you!
 
Hi, great instructions, and it worked well for me. 2:40 to run StarEXterminator on a 4kx4k image before, now 9 seconds!
i9-10980XE CPU, RTX3090 GPU.
 
Does anyone have recent experience with Amazon's GPU-enabled instances? The following instance types, in decreasing rough order of capabilities, have NVIDIA GPUs:

P3: NVIDIA A100 Tensor Core GPUs
P3: NVIDIA Tesla V100
P2: NVIDIA K80 GPU
G5: NVIDIA A10G Tensor Core GPUs
G5g: NVIDIA T4G Tensor Core
G4dn: NVIDIA T4 Tensor Core GPU
G3: NVIDIA Tesla M60 GPU

I'm planning to do the WBPP calibration as well as the *XTerminator processes on EC2, then do the final few steps on my local PI instance (which does have a GPU, but an older 1080, sometimes it chokes).

Thanks.
 
Last edited:
razvan said:
Does anyone have recent experience with Amazon's GPU-enabled instances? The following instance types, in decreasing rough order of capabilities, have NVIDIA GPUs:

P3: NVIDIA A100 Tensor Core GPUs
P3: NVIDIA Tesla V100
P2: NVIDIA K80 GPU
G5: NVIDIA A10G Tensor Core GPUs
G5g: NVIDIA T4G Tensor Core
G4dn: NVIDIA T4 Tensor Core GPU
G3: NVIDIA Tesla M60 GPU

I'm planning to do the WBPP calibration as well as the *XTerminator processes on EC2, then do the final few steps on my local PI instance (which does have a GPU, but an older 1080, sometimes it chokes).

Thanks.
Click to expand...
I am surprised the K80 is that high! You can pick up the 24gb K80 all day under 150 dollars.
 
That's my fault, I listed the GPUs in the inverse order of the EC2 "generations", but should have looked at the individual characteristics, they are not always in sync.

Well, in the meantime I made a few tests using the same Ubuntu 22.04 image with different instance types. The results were surprisingly underwhelming, particularly that the GPUs are not shared. Below I'm referring strictly to the single GPU performance, I have not measured anything else.

It took a g5.4xlarge instance with an A10G (9216 cores, 1.3GHz clock) to obtain the same times as my home GTX 1080 (2560 cores, 1.6GHz clock). A g4dn.4xlarge instance with a Tesla T4 (2560 cores, 0.6GHz clock) was at least 2x slower. So the clock frequency seems to have a larger impact than the number of cores. All tests were done by running BlurXTerminator on the same image from an ASI2600MC camera.
 
With my new Ryzen 7950 StarX took 8 minutes on a 5000x9000 image and 20s with CUDA!!!!!!!!!!!!!!!!!!!!!!

Tried 12.2 but had to go back to 11.8 but with the new DLLs.

This is a 3070 ti for gaming.... Microsoft Flight Simulator... you should see the graphics....
 
Nice job. Thank you so much. Worked like a charm in the first try. Impressive computing now.
 
You must log in or register to reply here.
Back
Top