Up-to-date instructions for enabling GPU acceleration

[RT--]

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

 

[BobT]

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

Thanks a million, I’ve just followed those instructions and what was taking StarXTerminator nearly six minutes it’s now doing in 30 seconds.

 

[RT--]

Thanks a million, I’ve just followed those instructions and what was taking StarXTerminator nearly six minutes it’s now doing in 30 seconds.

Nice! Which GPU do you have?

 

[BobT]

Nice! Which GPU do you have?

It’s a GeForce GTX1660 Ti

 

[jeffweiss9]

RT-
I followed your instructions above on instaling CUDA for PixInsight on my Dell Vostra 15-7510 with two display adapters: Intel UHD Graphics and NVIDIA GeForce RTX3050 Laptop GPU, assuming it would use the second GPU. I get very good times for CUDA BXT:24s, NXT: 12s, StarXT: 20s but the attachment shows that the Windows 10 Task Manager Performance Window, snapped while BXT was running, does not show any CUDA window as you show in your testing instructions. It shows instead that the two GPU's with more usage on GPU 0 (Intel UHD) than on the GPU 1 (NVIDIA GeForce RTZ3050 Laptop GPU). I am getting very fast times but its not clear to me that CUDA was even installed. But these times are so fast on this machine, its not clear that I'd get much benefit from CUDA if it is not running.
Any advice?
Jeff

 

[ktoliver]

Thanks so much for this. My time for processing a frame in Starnet2 just dropped from 8 minutes to 25 seconds!

 

[RT--]

RT-
I followed your instructions above on instaling CUDA for PixInsight on my Dell Vostra 15-7510 with two display adapters: Intel UHD Graphics and NVIDIA GeForce RTX3050 Laptop GPU, assuming it would use the second GPU. I get very good times for CUDA BXT:24s, NXT: 12s, StarXT: 20s but the attachment shows that the Windows 10 Task Manager Performance Window, snapped while BXT was running, does not show any CUDA window as you show in your testing instructions. It shows instead that the two GPU's with more usage on GPU 0 (Intel UHD) than on the GPU 1 (NVIDIA GeForce RTZ3050 Laptop GPU). I am getting very fast times but its not clear to me that CUDA was even installed. But these times are so fast on this machine, its not clear that I'd get much benefit from CUDA if it is not running.
Any advice?
Jeff

Those times indicate CUDA is working on your system. Without CUDA it would take a few minutes, and the CPU usage would be sitting at 100% the whole time.
The Cuda graph is not visible by default, you can select it from the dropdown by clicking 'Video encode'.
On some systems the Cuda graph is not available at all. I have not seen that myself but I have read that it is caused by hardware-accelerated gpu scheduling. You can try disabling it by searching the Start Menu for "Graphics settings"; there you'll find a toggle to turn off Hardware-accelerated GPU scheduling. Then reboot, and see if you get more graphs to choose from. Personally I've never had that setting turned on.

 

[jeffweiss9]

Thanks, RT-
Glad to confirm CUDA is working. I'll play around with trying to get the CUDA graph to display.
CS/Jeff

 

[fernandoyang]

Thank you very much for your tutorial.

 

[DaveSD]

RT - Thanks for the updated guide, easy to follow and worked perfectly!

I am setting up a new workstation. On a test case, Starnet2 ran in 8 seconds vice 191 seconds on the i7 based laptop this is replacing (~24X).

 

[astronut892010]

I seem to be having a problem compared to others. I followed your instructions step-by-step. There is no CUDA graph, but the 3D performance goes to 100 percent and memory usage increases. But SXT has no improvement. It still fluctuates between 1-2 min using the bare bones lite nonoise option. I am testing it on my gaming laptop. Here are my specs for the computer:

Processor Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz 2.21 GHz
Installed RAM 16.0 GB (15.9 GB usable)
System type 64-bit operating system, x64-based processor
Pen and touch No pen or touch input is available for this display

Graphics card NVIDIA GeoForce GTX 1050 Ti

Any suggestions for a direction to troubleshoot?

Jay Landis

 

[RT--]

What size images are you running it on?
Have you tried if StarNet2 shows any improvement?

 

[Suwat Panitkullawat]

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

Thanks for your great instructions. Did you test the latest cuDNN8.8.0 for CUDA11.8?

 

[oprache]

I have a Win11 machine with a Ryzen CPU and an RTX 2070. No glitch during installation but had to turn the Hardware Acceleration scheduler off to get to see the Cuda graph. However, it stays hopelessly stuck at 0%. PI does not crash. The Nvidia driver is the most recent. I re-entered the environment variables, checked all the dlls are where they need to be. Only difference with your instruction is that CUDA is now at 12.0 and, since I have only one machine, I downloaded the non-networked package. Any hints as to where I should look?
Olivier

 

[DaveSD]

Use the exact versions in the guide

 

[oprache]

Use the exact versions in the guide

Thank you Dave, I did that a little while ago and with CUDA 11.8 (the network version, still available from Nvidia) and the associated cuDNN: same results, the GPU is not being used. Unless it is a Windows 11 related issue, I do not know what else to try. Did soft and hard reboot, to no avail.

 

[DaveSD]

Not that I am an expert, but I'm sure it isn't a Windows 11 issue, it is compatible. I recommend you start from the beginning of RT's guide and redo the entire installation with new downloads. You must be either missing a step or using an incorrect version file. Good luck.

 

[RT--]

Thanks for your great instructions. Did you test the latest cuDNN8.8.0 for CUDA11.8

I did test it briefly but for me it was slower than the previous version, and it also drops support for the older cards, so my recommendation stays with 8.7.0.
RTX4000 series users might want to test it, though - maybe there's some extra performance for the latest generation.

CUDA 12 didn't work at all, so either my install method doesn't work with it, or it's incompatible with the tensorflow version I was using.
Besides, nVidia recommends CUDA 12 for the H100 GPU only, and says all the others get the best performance with 11.8.

When I have time, I'll do some more testing.

 

[umasscrew39]

I am installing this now but ran into one problem. I am trying to download
zLib DLL x64 1.2.3 (Right click > Save as...) but on my Windows 10 computer it downloads as unconfirmed. When I click on it, it wants to know what to open it with. It downloads just fine on my Mac. What am I doing wrong?

Bruce

 

[RT--]

I am installing this now but ran into one problem. I am trying to download
zLib DLL x64 1.2.3 (Right click > Save as...) but on my Windows 10 computer it downloads as unconfirmed. When I click on it, it wants to know what to open it with. It downloads just fine on my Mac. What am I doing wrong?

Bruce

Current browsers will block all downloads from sites that are not https. That's why you need to right click the link, and select 'save as'. Some browsers might then show a popup asking you to confirm the download. Try again, and read the message in the downloads window.