Error when attempting to render with Teslas

Hello there. I am in the process of building a classic workstation for funsies (circa 2011) and managed to get my hands on a couple of Teslas before everything went belly up.

I first put a K4000 into an older PC (Asus mini itx, i5 3570, 16GB DDR3 memory, and W7) to test it and installed the 321 Maximus driver with no issue during any tests. I pulled out the card and replaced it with a c2050 3GB which windows recognised immediately and ran the same battery of tests with no issue in Cinebench r15 and Blender Cycles. Despite running close to the 3GB VRAM limit I was able to render the Blenderman benchmark for 15 minutes before I canceled out.

I then got my hands on a c2075 which windows did recognise in the device manager but I could not open the Nvidia control panel as I got the “no NVidia card plugged in” error. Furthermore Windows would only run with the basic VGA driver, forcing me to run Blender on the IGPU while I would get the following error when attempting to render any scene in Blender more complex than the Koro demo with the Tesla:

CUDA error: Invalid value in cuTexRefSetAddress(NULL, texref, cuda_device_ptr(mem.device_pointer), size), line 750

I could duplicate the 2 BMWs from the GPU benchmark dozens of times until I got the normal out of memory error while preview rendering but attempting to enter render preview or execute a proper render with even half the number of cars I would get the aforementioned CUDA error. Putting the c2050 back in it behaved exactly as the c2075 but unlike the c2075 Windows used the proper Nvidia drivers and I could access its control panel. Still cannot render anything complex like the Blenderman demo which I could do so before.

Playing around with Tesla and Maximus drivers, uninstalling them with DDU, and trying the cards in different configurations on one of my main Workstations could not alleviate the issue. My K4000 has no trouble rendering anything that will fit on the VRAM while the two Teslas are throwing fits.

One other thing, the c2050 is properly reported by both GPUz and Afterburner while the c2075 throws up a bunch of zeros and unknowns and Afterburner does not even see it. This led me to believe that it was a fake card but the PCB checks out while the card, when it performs, does exactly as should be expected on paper. Including being able to saturate the entire VRAM (assumed from the memory usage reported in the render preview interface) in render preview mode.

Is the card defective, did it somehow affect the c2050 when I ran it in the same PC together with the Quadro driving display, is this a buggered setting in Blender, and is there any real way to test the c2075 to make absolutely that it is the card and not something else that I did wrong?

Thanks!

P.S. I have been using a single DVI to HDMI cord for all of the cards. Could this maybe affected them in some way and if so, why not the Quadro?

EDIT: I forgot to mention that this error persists across everything from version 2.79 to 2.74 (as I have read this a solution to similar but still different problems) before I gave up trying different versions. I have not tried the 2.8 build yet. I think I will give that a go quickly.

Hi.
You try Blender 2.79 from master/buildbot:

Download those labeled with “2.79” and “official”

Unfortunately I do not have the CUDA selection option available in the user settings (there is literally nothing there). 2.8 CTDs every time I try to open it as well so I cannot test either of them.

EDIT: I (poorly) remember a long time ago reading that the Fermi cards were actually limited in how much geometry, textures, etc, each could occupy in the memory such as 1.5GB geometry and 2GB textures but was this not eventually fixed?

It may be a bios setting that handles larger addressing.
used to be an issue in servers so maybe this link will help.


basically states you need to change an I/O mapping switch in your bios to allow for PCI memory addressing over 4GB.