Topaz 5.1.2 Fails to Utilize NVIDIA A40 GPU Despite Proper CUDA 12.4 Configuration

Video AI Bugs and Issues
1

Software and Environment:

  • Topaz Version: 5.1.2
  • GPU Model: NVIDIA A40
  • CUDA Version: 12.4
  • Operating System: win10

Problem Description: When conducting image processing tasks with Topaz 5.1.2, the application fails to leverage the NVIDIA A40 GPU, even after explicitly selecting GPU usage within the software settings and successfully installing a compatible NVIDIA driver and CUDA 12.4. During benchmark tests, GPU utilization monitored via GPU-Z remains at 0%, indicating that GPU resources are not being effectively harnessed, which significantly lags performance expectations.

image

image

This is benchmark info [NOT FINISH]:

Topaz Video AI  v5.1.2
System Information
OS: Windows v10.21
CPU: Intel(R) Xeon(R) Platinum 8462Y+  255.7 GB
GPU: NVIDIA A40  5.1209e+10 GB
GPU: NVIDIA A40  5.1209e+10 GB
Processing Settings
device: 0 vram: 1 instances: 0
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	03.49 fps 	2X: 	02.33 fps 	4X: 	00.90 fps 	
Iris		1X: 	04.45 fps 	2X: 	02.01 fps 	4X: 	00.68 fps 	
Proteus		1X: 	04.06 fps 	2X: 	03.00 fps 	4X: 	  ...    fps 	
Gaia		1X: 	   -     fps 	2X: 	   -     fps 	4X: 	   -     fps 	
Nyx		1X: 	   -     fps 	2X: 	   -     fps 	
Nyx Fast		1X: 	   -     fps 	
4X Slowmo		Apollo: 	   -     fps 	APFast: 	   -     fps 	Chronos: 	   -     fps 	CHFast: 	   -     fps 	
16X Slowmo		Aion: 	   -     fps
1 Like
2

Try to use it via the CLI.

As far as i know the A40 is headless and does need a other GPU for display because of DirectML. (seems like i’m wrong, the A40 has 3x DP.)

Or i’m wrong and you should reach out for support on the TL mainsite.

3

btw

Which driver did you install?

Install the driver for the Professional cards not the server ones.

4

Some things in that GPU-Z screenshot don’t look right. It doesn’t look like it’s reporting some of the stats at all. It appears that some system/driver issues need to be sorted out first.

Definitely contact support directly for help with data centre gpu configurations.

6

I solv this issue, A40 need change the the compute mode to graphics mode and change to the WDDM mode

1 Like
7

Another proof that GPU are used as renderer GPU and not compute mode ?

Here we aren’t in games right ? So why programming them as if they were ? Also sounds as CUDA is not used

8

This is probably caused by the way the app GUI detects GPUs using DirectX. Model processing does use Tensor cores or CUDA cores depending on the model, but I will check with the dev team to see if compute mode on cards like the A40 could be supported.

1 Like
9

Thank you very much for paying attention to this, because owning stuff from Pascal to Ada Lovelace, I’ve so much difference in behaviours:

  • GPU used as renderer, no CUDA (CUDA Do exist starting as Pascal)
  • CUDA loads on some gen, but not all, strange things …
  • Tensors use effective, but other shaders not used at all ?

I also think to AMD users, because performance are by far worst than Nvidia

  • OpenCL used by anyway, or again GPU used as renderer ?

Starting to Navi+ there is RDNA, that can be used as compute engines (very efficently)

Regards

10

Starting with Tesla in 2007, Pascal is 2016.

CUDA is a Programming language, as OpenCL is too.

Most AI Images software (from big vendors) is making use of DirectML because you can use it on Nvidia, AMD and Intel and not just one vendor.

And the discussion about whether CUDA wouldn’t be better is already almost 4 years old.

11

Sorry for not having read the last 4 years of threads …

12

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.