GTX Titan double precision performance under OSX

[Michiel Otten]

Hi,

I was wondering what the double precision performance of an NVIDIA GTX Titan is under OSX. As posted in several places (e.g., by Tutor in post #564 https://forums.macrumors.com/threads/1333421/) you have to enable the CUDA double precision option in the NVIDIA Control Panel under Windows. Doing so will increase the FP64 performace from ~230 Gflops to ~1500 Gflops.

I need double precision accuracy for the work I do and ~230 Gflops by no means bad is only twice as fast as my CPU and not really worth $1000 but ~1500 Gflops well now we are talking

So it would be a big help to me if somebody who has a Titan installed in his Mac could post a screen shot of CUDA-Z under OSX or even better run testing_dgemm which is part of the NVIDIA CUDA drivers.

A big thanks in advance!
Michiel.

 

[Tutor]

Hi,

I was wondering what the double precision performance of an NVIDIA GTX Titan is under OSX. As posted in several places (e.g., by Tutor in post #564 https://forums.macrumors.com/threads/1333421/) you have to enable the CUDA double precision option in the NVIDIA Control Panel under Windows. Doing so will increase the FP64 performace from ~230 Gflops to ~1500 Gflops.

I need double precision accuracy for the work I do and ~230 Gflops by no means bad is only twice as fast as my CPU and not really worth $1000 but ~1500 Gflops well now we are talking

So it would be a big help to me if somebody who has a Titan installed in his Mac could post a screen shot of CUDA-Z under OSX or even better run testing_dgemm which is part of the NVIDIA CUDA drivers.

A big thanks in advance!
Michiel.

Michiel,

Under Windows and using EVGA Precision X, that 1.5 Tflops can easily and safely be increased to 2 Tflops and beyond.

Another source of potentially useful intel would be from folks running the GTX 680 Mac Edition who can tell you whether there is a Mac Control Panel and what does it allow and whether there is a Mac utility akin to Precision X to modify the thermal and power settings and the offsets for memory and GPU clocks. Such software might also work with Titan.

 

[Michiel Otten]

Michiel,

Under Windows and using EVGA Precision X, that 1.5 Tflops can easily and safely be increased to 2 Tflops and beyond.

Another source of potentially useful intel would be from folks running the GTX 680 Mac Edition who can tell you whether there is a Mac Control Panel and what does it allow and whether there is a Mac utility akin to Precision X to modify the thermal and power settings and the offsets for memory and GPU clocks. Such software might also work with Titan.

Tutor,
thanks for the reply I do have access to a Mac Pro with a GTX 680 and I could not find anything similar to the Windows Control Panel on the Mac (this Mac Pro had the CUDA 5.0 drivers installed).

I know that on the Mac you can overclock the GPU using the method described here by Dr Stealth's thread on the 4GB GTX 680 cards: https://forums.macrumors.com/threads/1603260/

The GTX 680 is an amazing card but unfortunately not in FP64 mode because it only has 1/24 its FP32 performance which makes the card just as fast as my Mac Pro 5,1 3,33GHz hex core (around ~100 Gflops)

Further I realised that the dgemm test I mentioned is actually part of MAGMA and not the standard CUDA drivers of NVIDIA. MAGMA is a very cool library for people needing LAPACK/BLAS functionality on a GPU (http://icl.cs.utk.edu/magma/)

Cheers,
Michiel.

 

[Tutor]

Tutor,
thanks for the reply I do have access to a Mac Pro with a GTX 680 and I could not find anything similar to the Windows Control Panel on the Mac (this Mac Pro had the CUDA 5.0 drivers installed).

I know that on the Mac you can overclock the GPU using the method described here by Dr Stealth's thread on the 4GB GTX 680 cards: https://forums.macrumors.com/threads/1603260/

The GTX 680 is an amazing card but unfortunately not in FP64 mode because it only has 1/24 its FP32 performance which makes the card just as fast as my Mac Pro 5,1 3,33GHz hex core (around ~100 Gflops)

Further I realised that the dgemm test I mentioned is actually part of MAGMA and not the standard CUDA drivers of NVIDIA. MAGMA is a very cool library for people needing LAPACK/BLAS functionality on a GPU (http://icl.cs.utk.edu/magma/)

Cheers,
Michiel.

Michiel,

Thanks for introducing me to dgemm test and to Dr. Stealth's enlightening thread. I'd missed DS's chronicle. There's more to it than meets the eye. The "more" is what I'm always after.