Problem Initializing GPUs in "Select Open CL Devices"

[dpmitchell]

Hello,

I just completed a new build with the following specs:

Asus Rampage III Formula
Intel i7 950
12 GB DDR3 1333
Nvidia Quadro 4000 (2GB GDDR5)
Nvidia GTX 460 (1GB GDDR5)

I have installed Vray RT 1.9 beta (Sept 23 release).

When I go into the "Select Open CL Devices" in the Vray RT startup menu, the box shows both the Quadro 4000 and the GTX 460. Regardless of which device I check, when I hit "Set Devices" I get the following error message: "Failed to update the open CL devices to use.

GPU-Z confirm that both Open CL, CUDA, Physx and Direct Compute 5.0 is available on both devices.

When I open max and pull up the Vray RT active shade properties, it gives me the option to select CPU or single kernel OpenCL, however, if attempt to render an active shade window with the Open CL option, nothing appears in the window. If I select CPU, it renders as usual Vray RT with CPU.

I assume this may be a driver problem. This is the first time I've used multiple GPU's in one build, but even when I only have one GPU in the machine, I get the same error message. Also, keep in mind that the GTX 460 was not added to Nvidia's compatible Geforce drivers list until Geforce release 258.96. The Quadro 4000 wasn't updated on Nvidia's Quadro drivers until Quadro release 258.96. The systems seems to want the same driver installed on both cards, which was troublesome. I've even used a modified Quadro 258.96 that will install on the GTX 460 as well as the Quadro 4000, and no cigar there either (although the GTX 460's Cinebench 11.5 Open GL and Max viewport performance did seem a bit improved with the Quadro driver when I did some solo testing).

I know this mixing of Quadro for viewports and Geforce GTX 400 series for rendering combination is possible as someone did an article comparing a Quadro + Tesla combination to a Quadro + GTX 480 combination (interestingly, the GTX outperformed the Tesla on pure rendering numbers, but memory limitations and increased heat were the obvious concerns. The 480's issues might be alleviated by the GTX 580 which runs much cooler due to the refined GF110 fermi chip and 1.5 GB memory).

Any help would be much appreciated!

[vlado]

If you install the nVidia GPU Computing SDK, are you able to run their OpenCL samples?

Best regards,
Vlado

[dpmitchell]

Thanks for the response Vlado. I installed the GPU Computing SDK, and I was unable to run any of the CUDA C or Open CL Samples. I was able to run all of the Direct Compute Samples, but it only allowed me to run them on the Quadro 4000 card. The GTX 460 was not listed in the drop down window in those tests.

Most of the Cuda C and Open CL samples briefly opened and closed a cmd prompt typ window when I chose "Run". Some of them, such as the OpenCl Post Process Open GL-Rendered Image test returned the following error:

!!! Error # -999 <unspecified error> at line 543, in file .\oclPostprocessGL.cpp !!!

Starting Cleanup

Failed

oclPostProcessGL.exe Exiting"

The OpenCL Simple Texture 3D returned this error:

!!! Error #0 at line 236, in file .oclSimpleTexture3D.cpp!!!

etc.

Thanks for your help! I hope we can get this figured out. I can't do a lot of manual tweakingt to the system at the mometn because I'm logged in remotely, but if this is a driver fix, then I hope I can get it resolved.

[vlado]

What do you use to log in remotely? Using Remote Desktop will not work, as it will use a special software display driver which does not support access to the GPU. Using LogMeIn works though.

Best regards,
Vlado

[dpmitchell]

Yes, I'm using remote desktop right now. I will have to try running the GPU computing sdk again when I get home. But all of the problems I was experiencing in the initial post were happening when I was sitting in front of the actual machine, not on remote desktop. I will post an update later. Thanks for the info...I didn't know about LogMeIn. I'll have to check that out.

[dpmitchell]

Vlado,

I just logged in using LogMeIn (neat, but slow), and ran the GPU Computing SDK again. This time, all of the Open CL tests seemed to run fine. Is there something I should be looking for in particular?

Thanks!

[vlado]

Ok then; you can try to run the OpenCL device selector with Administrative privileges (right-click the shortcut and then Run as Administrator...), it tries to set a system environment variable, so if you don't have administrative privileges, it will fail.

Best regards,
Vlado

[dpmitchell]

Wow I feel stupid. I can't believe I didn't try that already. Ok good deal! The box stated that both devices were found and updated, and the changes would apply next time the Vray RT render server runs.

I'll test this 460 a little, but I'm replacing it with the 2GB version next week. I just realized they made one a couple days ago. Nvidia's specs show that the 460 comes in only 768 and 1GB. However, I found at least 2 aftermarket vendors that are producing a 460 with 2 GB. That actually makes it the best of the Geforce GTX fermi line for my purposes so I don't end up with a bottle neck on the data set being that my Quadro 4000 is running 2GB. No other Geforce GTX fermi card has a 2GB model.

Also, I can confirm that the 460's GF104 fermi chip runs quite a bit cooler than it's predecessor the GF100. The Quadro 4000 uses the GF100 fermi chip (as well as the GTX 470 and 480), and the 460 idles at about 19-23 C cooler than the Quadro, despite the fact that the Quadro had a third of it's CUDA cores disabled, its memory bandwidth reduced, and clock frequency reduced in order to draw less power and improve stability.

The early reports are that the new GF110 fermi (found in GTX 580) actually has improved thermal management beyond the GF104. Runs cool as a cucumber. Unfortunately the largest currently available memory on the GTX 580 is 1.5 GB. But I bet it won't be long before Zotac, EVGA, Palit, etc pops out a 3GB version which could serve the dual purpose of being a budget Tesla and the newest unnecessary "gotta have it" item for gamers. Market diversification

Now what will Nvidia do to cripple it? LOL