Correct Driver for CUDA 6.5 Support?

I have to call upon the mothership–aka NVIDIA–of PGI with this question. Namely, what driver is needed to run with CUDA 6.5? For example, I go here:

http://www.nvidia.com/Download/index.aspx?lang=en-us

and select my Tesla[1] and then ask for the Linux x86_64 driver and it gives me:

TESLA DRIVER FOR LINUX X64
 
Version:	331.89
Release Date:	2014.7.23
Operating System:	Linux 64-bit
Language:	English (US)
File Size:	60.00 MB

Okay. That’s the latest, but when I try it out with PGI 14.9 pgaccelinfo:

(169) $ pgaccelinfo

CUDA Driver Version:           6000
NVRM version:                  NVIDIA UNIX x86_64 Kernel Module  331.89  Tue Jul  1 13:30:18 PDT 2014

That definitely isn’t 6050, but then NVIDIA says the correct driver is 331.89, so who am I to judge. However, when I try to run 6.5 code I get the expected error:

pgfortran -fast -r4 -Mextend -Mpreprocess -Ktrap=fp -Kieee -Minfo=accel -Mcuda=nofma,6.5,cc20,ptxinfo -Mcuda=maxregcount:72 -DNITERS=6 -DGPU_PRECISION=8 driver.o sorad.o timingutils.o statutils.o soradconstants.o radconstants.o gettau.o -o sorad.GPU.exe
make[1]: Leaving directory `/home/mathomp4/G21p3-Benchmark-Sorad'
mv sorad.GPU.exe GPU/
cd GPU; ./sorad.GPU.exe; cd ..
 system_clock rate     10000000.00000000     
 cudaSetDevice failed: 
 CUDA driver version is insufficient for CUDA runtime version

Does anyone here or at NVIDIA know the magic words to use on the driver page to get the correct one? I tried asking for a K40-era driver (all four versions) but it still only points to the 331.89.

Now, if I select a GeForce TITAN Z, the page returns the 340.32 driver and according to the README that should support the S2050, K40, etc., but…well, I’d prefer to use the one NVIDIA wants me to from the official download page. I learned long ago that sometimes the latest display driver often isn’t the best for Tesla cards.

Thanks,
Matt

[1] Well, due to a coding bug in the webpage, I have to select a different Tesla. I have an S2050, but the fine folks that run the page decided to not to add the S2050 as a viable chip for the 331.89 driver even though the 331.89 driver README says it is supported. I figured if the C2050 is, the S2050 is so I selected that.

Hi TheMatt,

The CUDA 6.5 driver comes with the CUDA 6.5 Toolkit. Hope this helps.

Cheers,
Kyle

Kyle,

Huh. You are right. I’m guessing the admins here didn’t download the toolkit because I usually tell them not to since PGI includes everything you need for CUDA Fortran. Well, everything but the NVIDIA driver.

I’m still confused as to why the driver in the toolkit (340.29) doesn’t match the driver on the NVIDIA website. Shouldn’t they be consistent?

Matt

Hey Matt,

The driver on the website should indeed be equal to or newer than that in the toolkit. This is the first time I’ve ever seen that it isn’t. I’m sure it’ll get updated soon.

-Kyle

Hi Matt,

The Tesla drivers undergo more rigorous testing between the time they are released in the CUDA developer’s kit and the time they are released on the main driver page. Hence, the lag time. You can go to the UNIX driver archive page if you want to download just the latest driver without downloading the full toolkit. http://www.nvidia.com/object/unix.html

  • Mat