340.52 driver breaks OpenCL support for pre-FERMI (CC 1.x) cards

Accelerated Computing CUDA CUDA Programming and Performance
1

  1. After all those years passed there is no separate OpenCL subforum here. Nice to see strong nVidia’s commitment to open standarts, indeed…

  2. Our testers established that with driver 340.52 pre-FERMI cards generate wrong results with OpenCL app. With 337.50 (for example) app produces correct results on the same hardware.
    Can we expect fix for this issue in upcoming driver update or nVidia’s OpenCL support is ruined completely?

2

I can’t speak to the specific issue you are referring to, as it is unclear from your report. However, in general, for both CUDA and OpenCL, and in fact general usage (e.g. graphics), support for pre-FERMI is going away in the latest NVIDIA drivers, in the near future. CUDA toolkit 6.5 deprecates support for these devices, and in the near future the driver will drop support.

For existing pre-FERMI hardware, you should probably remain on previous drivers. For future support, you should probably migrate to newer hardware.

3

Thank you for answer.
Drop of support is not good news but expectable at some point of time.
The real issue is that 340.52 driver simulates work. It doesn’t throw error, app under it works “ok”, just generates incorrect results.
Such behavior is unacceptable. If no support for particular device - don’t detect it or throuw error if some deprecated call is done. But not silently provide wrong computations.
I filed bug report with CUDA developer program.
Bug ID is #1554016.

4

I’ve been able to identify places using my Quadro FX 3800M, where [R340 341.05 Cuda6.5] will FAIL some of NVIDIA’s own OpenCL SDK examples, whereas [R337 337.88 Cuda6.0] runs the examples fine… on Windows 8.1 x64.

And, the errors are DATA errors! Something’s definitely broken in the drivers! We need NVIDIA to fix this!

[url]https://developer.nvidia.com/opencl[/url]
The examples that are giving me data verification errors with 341.05, are:

  • oclFDTD3d
  • oclDXTCompression
  • oclQuasirandomGenerator
  • oclConvolutionSeparable

Can anyone confirm the data failures on those examples?
For me, those 4 examples fail on all R340 drivers: 340.43, 340.52, 340.66, 340.84, 341.05.
They worked fine on R337 drivers: 337.88

We need NVIDIA to fix this!

– Jacob

5

Anyone know if this issue also present on a Tesla C1060 for example? Or is it only GeForce cards that are affected?

6

It seems that removal of support for sm_1x GPUs in the Windows drivers is either imminent, or has happened already:

[url]Error | NVIDIA

How the driver generation numbers mentioned in the above end-of-life notice relate to the specific driver version numbers mentioned above I do not know, as NVIDIA’s driver numbering scheme has always been a bit of a mystery to me.

7

For reference, I have NVIDIA Bug 1574543 open presently.

NVIDIA vows to maintain the R340 branch of drivers for these pre-Fermi devices, until April 2016, and hopefully they will fix the OpenCL issues that I’ve documented earlier in this thread.

Regards,
Jacob

8

I have confirmed that the recently-released R340 driver version 341.44, fixes the R340 OpenCL silent data errors that we were having on pre-Fermi hardware running on prior R340 drivers.

Regards,
Jacob

9

http://www.nvidia.com/download/driverResults.aspx/82465/en-us

http://www.nvidia.com/download/driverResults.aspx/82467/en-us

Direct links to 341.44 WHQL driver.