GPU Computing

cancel
Showing results for 
Search instead for 
Did you mean: 

what cuda version do I need?

Solved!
Go to solution

I noticed that the latest version of cuda (10.1) gives false results when computing FFT.
A possible solution is to install cuda version 4.1. (maybe a newer version also works but versions >= 9 suffer from the same problem)

 

The latest version of ni GPU analysis toolkit (V15.0.0.6) has been released in 2016. (via VIMP)

 

The more recent (relevant) posts on this forum are from 2017.

 

Is this toolkit still being maintained or is slowly dying?

 

If there is still a big interest in this toolkit, can someone provide me with a procedure to install the more recent cuda version (with a working fft) that work with the more recent version of labview?

 

 

 

 

0 Kudos
Message 1 of 4
(4,660 Views)

Could you please go in more detail, regarding:


@sambuls2 wrote:

I noticed that the latest version of cuda (10.1) gives false results when computing FFT.

 

what hardware  are you using?

what labview version are you using?

 

where's your .vi - how did you compute the FFT?

are you referring to the  FFT example from NI?

 

 

 

0 Kudos
Message 2 of 4
(4,649 Views)

My specs:

GPU Device: Quadro M6000

CUDA Version: 10.1

Cuda Display Driver: 10.1

Labview version: V19.0 (32-bit)

NI GPU Analysis Toolkit 32-bit

The result of 'Get Version VI' for FFT, when using CUDA 10.1, gives a CUFFT version of: 0.0 (?)

 

I added two printscreens of the front panel when running the Multi-channel FFT example.

You can see that the result calculated with CUDA version 4.1 is different (correct) than the result calculated with CUDA version 10.1 (incorrect).

Even when distilling the multi-channel FFT example to a more simpler version, the result from 10.1 is still incorrect.

 

From my recent experience, CUDA version 6.5 is also working.

 

Do you have any advice?

 

Download All
Message 3 of 4
(4,641 Views)
Solution
Accepted by topic author sambuls2

@sambuls2 wrote:

Do you have any advice?

 


Thank you for pointing out your specs,

this appears to be related to this:

 

https://knowledge.ni.com/KnowledgeArticleDetails?id=kA00Z0000019QTVSA2&l=de-DE

The CUDA Toolkit driver that the GPU Analysis library depends on removed support for 32-bit Windows in version 7.0. As a result, you cannot use LabVIEW 32-bit with newer versions of the CUDA driver. If you'd like to use the LabVIEW GPU Analysis Toolkit for FFT, you have the following options:

  • Use LabVIEW 64-bit on a Windows 64-bit system
  • Revert to an older version of the CUDA driver (older than 7.0)

 

So, using CUDA 6.4 is a valid work around

 

 

 

 


@sambuls2 wrote:

Is this toolkit still being maintained or is slowly dying?

 

If there is still a big interest in this toolkit, can someone provide me with a procedure to install the more recent cuda version (with a working fft) that work with the more recent version of labview?

I suppose, we can only expect  future releases for LabView x64.

 

The latest version of the GPU Toolkit came out with LabView 2014 (x64)

https://www.ni.com/de-de/support/downloads/software-products/download.labview-gpu-analysis-toolkit.h...

 

But: I remember back in 2014 this Toolkit was not a public toolkit, as it is now.

You had to pay around 1000 $ or € to use the "LabVIEW GPU Analysis Toolkit"

 

So, why did NI make this toolkit available for public?

I suppose, they didn't sell many of copies of that toolkit...

 

 

 

 

 

Message 4 of 4
(4,620 Views)