How to do that?
Experimental TensorFlow GPU acceleration repository
- Thread starter rcroman
- Start date
Unfortunately, last I checked at least, DirectML had not been updated in at least a year, and was lacking support for operations used in RC Astro tools. Otherwise it would most likely be the right solution.
I've not tried ZLUDA – I don't have a machine I could test it on.
I've not tried ZLUDA – I don't have a machine I could test it on.
Hi. Thank you for your effort on this - it should certainly help a lot of people.
I hope to run with Cuda enabled under linux soon, but do not have time to try to make it work at present. I will be following your progress on this!
My current setup is a Lenovo TinkBook 15 G2 (i7-1165G) with on board MX450. Cuda acceleration sped up StarXTerminator from 10 minutes to 50 seconds with the MX450 running Tensorflow etc. As I already have this running under Windows, I will follow your advice, and not mess with it (well, that's what I _thought_!)
My main image processing machine crapped out (e5-2969, 64Gbyte RAM, GTX1660ti etc.) so I set about to repurpose the GPU and power supply. Ended up getting a handy eGPU (thunderbolt 4) adapter. FINALLY managed to get all the drivers etc. working. Last thing that helped was to install the latest Nvidia drivers in safe mode with the GTX1660ti already installed. Since then, it has worked very well so far. That said, the laptop now thinks it's got 3 GPU's (built in, MX450 and eGPU of GTX1660ti.)
The performance improvements are significant. Whilst I don't have Starnet installed, I do have StarXTerminator. As above, CPU only is 10 minutes which is rather long.... With MX450 Cuda (AI version 11 lite as MX450 only has 2Gbyte RAM) takes this down to 50 seconds - a huge increase in speed although using the AI 11 lite version.
With the GTX1660ti, this goes down to 20 seconds (same AI version - 11 lite). As the 1660ti has 6 Gbyte RAM, I can use the normal (not lite) AI version. This then takes 25 seconds, but at least I can run it.
I may be brave enough one day and try your scripts over the top of what I've got - would be interesting to see any differences. And am planning on doing a Linux install as well, and back to running PI under Linux, but really want to be sure I can get Cuda working there first.
Oh, yeah, have to disable the MX450 in order for the 1660ti to be used - small price to pay! And next time someone I know is 'upgrading their GPU', I might be able to pickup their dud 'old' one and put it to use...
Looking forward to seeing any updates.
I hope to run with Cuda enabled under linux soon, but do not have time to try to make it work at present. I will be following your progress on this!
My current setup is a Lenovo TinkBook 15 G2 (i7-1165G) with on board MX450. Cuda acceleration sped up StarXTerminator from 10 minutes to 50 seconds with the MX450 running Tensorflow etc. As I already have this running under Windows, I will follow your advice, and not mess with it (well, that's what I _thought_!)
My main image processing machine crapped out (e5-2969, 64Gbyte RAM, GTX1660ti etc.) so I set about to repurpose the GPU and power supply. Ended up getting a handy eGPU (thunderbolt 4) adapter. FINALLY managed to get all the drivers etc. working. Last thing that helped was to install the latest Nvidia drivers in safe mode with the GTX1660ti already installed. Since then, it has worked very well so far. That said, the laptop now thinks it's got 3 GPU's (built in, MX450 and eGPU of GTX1660ti.)
The performance improvements are significant. Whilst I don't have Starnet installed, I do have StarXTerminator. As above, CPU only is 10 minutes which is rather long.... With MX450 Cuda (AI version 11 lite as MX450 only has 2Gbyte RAM) takes this down to 50 seconds - a huge increase in speed although using the AI 11 lite version.
With the GTX1660ti, this goes down to 20 seconds (same AI version - 11 lite). As the 1660ti has 6 Gbyte RAM, I can use the normal (not lite) AI version. This then takes 25 seconds, but at least I can run it.
I may be brave enough one day and try your scripts over the top of what I've got - would be interesting to see any differences. And am planning on doing a Linux install as well, and back to running PI under Linux, but really want to be sure I can get Cuda working there first.
Oh, yeah, have to disable the MX450 in order for the 1660ti to be used - small price to pay! And next time someone I know is 'upgrading their GPU', I might be able to pickup their dud 'old' one and put it to use...
Looking forward to seeing any updates.
Last edited:
cloudbait
Well-known member
I think you're referring to a startup PI message? That has nothing to do with whether you have GPU acceleration, which is something that PI doesn't use at this point.
Correct... this only works with NVIDIA GPUs.
Yes the startup message.
I got 2 systems , one with nvidia (RX3060Ti card and 12th gen intel cpu) and an older one I store my images and usually process it (AMD 5-3600+ RX 6600XT) , but I don't think it's working on my AMD system. It's very slow, just like not having any graphics cards. I tried resetting/updating/reloading the repositories, but it won't work regardless. I mean I understand - 1st world problems - I could just work on the other machine of course, but that means I would have to migrate all my images and install pixinsight and re-check all the files to install, because I don't have anything astrophotography stuff installed on that system.
cloudbait
Well-known member
The GPU acceleration only impacts the three Croman tools, though. Any other differences you are seeing in performance between the two systems is related to other features of those systems. Not to their graphics cards.
Thanks! Beside the tools, interestingly the newer system isn't much faster than the older one. Running benchmarks the newer Intel+Nvidia system is ahead, but not by much really. Without seeing the benchmark scores, I wouldn't be able to tell. I actually prefer the AMD because it runs more silent and less heat and less power usage, but again, I don't want to go off topic here. The tools still work fine, just take a minute or 2 longer.
Just updated to build 1605 on RTX3060 windows PC. It removed, as per usual, BlurX, and Starnet2. I reinstalled both.
Tried both BlurX and Starnet2 on linear .xsif file (of galaxy NGC5754). GPU acceleration is gone. CPU cranks up to 90%-100%, just like it should when GPU acceleration is not enabled. I had backed up tensorflow file prior to updating PI, but I see it did not change the current file in PixInsight bin directory. I tried installing https://www.rc-astro.com/TensorFlow/PixInsight/GPU and realized it was already installed.
Many resets, power offs, complete removal (in control panel) and reinstall of PI build 1605. I also see GPU accel enabled in global preferences.
My only point here is that I'm so sick of this hacker crap. I have worked with computers and different OS since the 1980's for both business and personal use. The last 3 PI updates have all caused various problems.
I can follow directions. PixInsight is not reliable when updating the version and or build, and version/build updates are becoming more frequent.
As an end user, I must say that PI suffers extremely poor performance when executing updates. It makes me recall things like DOS 4.0...
P.S. I'll try GPU accel the old manual way after I get some sleep.
Tried both BlurX and Starnet2 on linear .xsif file (of galaxy NGC5754). GPU acceleration is gone. CPU cranks up to 90%-100%, just like it should when GPU acceleration is not enabled. I had backed up tensorflow file prior to updating PI, but I see it did not change the current file in PixInsight bin directory. I tried installing https://www.rc-astro.com/TensorFlow/PixInsight/GPU and realized it was already installed.
Many resets, power offs, complete removal (in control panel) and reinstall of PI build 1605. I also see GPU accel enabled in global preferences.
My only point here is that I'm so sick of this hacker crap. I have worked with computers and different OS since the 1980's for both business and personal use. The last 3 PI updates have all caused various problems.
I can follow directions. PixInsight is not reliable when updating the version and or build, and version/build updates are becoming more frequent.
As an end user, I must say that PI suffers extremely poor performance when executing updates. It makes me recall things like DOS 4.0...
P.S. I'll try GPU accel the old manual way after I get some sleep.
tom886
Well-known member
This doesn't sound like you really followed directions. When you update to a new build, Pixinsight installs a tensorflow.dll, but that's not the one you want. Because this tensorflow.dll is not for GPU acceleration.
So, no matter what, after an upgrade you have to replace the tensorflow.dll and overwrite the one in the Pixinsight directory.
Yeah, you can enable that or disable that as long as you want, Pixinsight doesn't support CUDA (yet). I believe it's written there.
GPU acceleration is only supported by the XT-tools and Starnet2.
Cheers
Tom
Ah, of course, I must have done <insert any term> wrong. I did install the backup tensorflow.dll, as one of the many iterations that I tried, in order to troubleshoot the root cause of the problem. I also did see that the tensorflow.dll was the exact filesize, time and date after the 1605 install as prior to the install. I purposely checked for that dll in program files\pixinsight\bin directory because I was aware of BlurX, starnet2 needing to be reinstalled when PI update removes them. I can't believe it's impossible to code a solution that retains any process and script plugins that are installed prior to PI update. That would be the engineering solution to this problem. I'll figure it out eventually. I do maintain the position that PI update is not a robust process.
Take Care,
Ralph
Russ,
Can the Photoshop version of StarXTerminator be accelerated by replacing the tensorflow.dll in C:\Program Files\Common Files\RC-Astro\StarXTerminator with the GPU version that was installed in C:\Program Files\PixInsight\bin with this update?
BTW, the update worked great with my RTX 3080 Ti.
Thanks,
Behyar
Can the Photoshop version of StarXTerminator be accelerated by replacing the tensorflow.dll in C:\Program Files\Common Files\RC-Astro\StarXTerminator with the GPU version that was installed in C:\Program Files\PixInsight\bin with this update?
BTW, the update worked great with my RTX 3080 Ti.
Thanks,
Behyar
sn2006gy
Well-known member
On Linux / Ubuntu I built libtensorflow for ROCm and these work on AMD GPUs now. https://rockchucksummit.com/pixinsight-gpu-acceleration-for-amd/
This would likely work if ALL of the DLLs included in the repository are copied to that directory:
Code:
cublas64_11.dll
cublasLt64_11.dll
cudart64_110.dll
cudnn64_8.dll
cudnn_adv_infer64_8.dll
cudnn_cnn_infer64_8.dll
cudnn_ops_infer64_8.dll
cufft64_10.dll
curand64_10.dll
cusolver64_11.dll
cusparse64_11.dll
nvrtc64_112_0.dll
tensorflow.dll
zlibwapi.dllCopying these into the folder did not improve the speed.
astro.lars
New member
Hi there,
I am currently using a Dell Mobile Precision 3591 (Core Ultra 7 155H, 32GB RAM, RTX A1000) (Mobile Workstation GPU) with Windows 11. The graphics card is not listed on the NVIDIA website, but is definitely NVIDIA Compute Capable and has NVIDIA Compute Capability v. 8.6 in my version.
This has already been confirmed in the official NVIDIA forum and the graphics card can perform calculations via CUDA. I also did a search in the Matlab forum, as such cards are often used for calculations, and also found that it is supported.
But no matter how I configure PixInsight for CUDA, it won't work... whether via the repository or via the manual variant ... no chance. I have also tried a newer TensorFlow version and a newer NVIDIA CUDA Toolbox...
On my tower with an older GTX 1070 it runs without problems, was ready to go after just 5-10 minutes and then calculated immediately via BlurX. However, Pixinsight only uses the CPU, neither the GPU nor the NPU. (I don't even know if NPUs are supported yet)
I have also used the NVIDIA system settings for PixInsight to specify that it should use the NVIDIA card for this application. CUDA acceleration is also activated in PixInsight. I have now uninstalled everything again and set it up again, it just doesn't want to work.
Does anyone have a similar card or even a similar device and got it to work?
Clear Skies
Lars
I am currently using a Dell Mobile Precision 3591 (Core Ultra 7 155H, 32GB RAM, RTX A1000) (Mobile Workstation GPU) with Windows 11. The graphics card is not listed on the NVIDIA website, but is definitely NVIDIA Compute Capable and has NVIDIA Compute Capability v. 8.6 in my version.
This has already been confirmed in the official NVIDIA forum and the graphics card can perform calculations via CUDA. I also did a search in the Matlab forum, as such cards are often used for calculations, and also found that it is supported.
But no matter how I configure PixInsight for CUDA, it won't work... whether via the repository or via the manual variant ... no chance. I have also tried a newer TensorFlow version and a newer NVIDIA CUDA Toolbox...
On my tower with an older GTX 1070 it runs without problems, was ready to go after just 5-10 minutes and then calculated immediately via BlurX. However, Pixinsight only uses the CPU, neither the GPU nor the NPU. (I don't even know if NPUs are supported yet)
I have also used the NVIDIA system settings for PixInsight to specify that it should use the NVIDIA card for this application. CUDA acceleration is also activated in PixInsight. I have now uninstalled everything again and set it up again, it just doesn't want to work.
Does anyone have a similar card or even a similar device and got it to work?
Clear Skies
Lars
astro.lars
New member
Okay, I downloaded the OpenHardwareMonitor and looked at the clock rates of the graphics card. Apparently it does use the graphics card. After the calculations, the GPU clock rate drops again. The Windows Task Manager just shows crap!