Geforce 1080ti vs Quadro P4000 for neural networks and deep learning

Which one is better (at the same price) for training shallow and deep networks in neural network toolbox (r2017b)?

Looking at their specs (11,340 gflops vs 5,304 gflops) the 1080ti seems the one with higher performance/cost ratio but I heard of some cuda issue with geforce line in neural network toolbox. Is 1080ti fully supported in r2017b release? Any other advice?

Thanks

[Walter Roberson]

The 1080 series was not fully supported in R2016b, but that has been fixed I believe.

Are you planning to use single precision or double precision? The two have significant double precision performance differences.

[Roberto]

I never changed the standard setting in matlab that i think is double precision. What's the problem with double precision and 1080 series?

Answer 1

[Joss Knight]

On the face of it the GTX 1080Ti is an all-round better performing card than the Quadro P4000 for deep learning applications, with better compute power and memory size. However, it will have considerably worse double precision performance. Deep learning is all in single precision, but if you are supplementing your deep learning with any standard computation, for instance systems modelling, data analysis or statistics, you will see performance on the 1080 barely much better than your CPU.

The 1080 doesn't support remote desktop which means that if you want to access the machine remotely, you'll find you can't use the GPU if you're using it for graphics as well. The current best-in-show for deep learning on the desktop is the Titan Xp.

[Walter Roberson]

There is a nice chart at https://www.microway.com/knowledge-center-articles/comparison-of-nvidia-geforce-gpus-and-nvidia-tesla-gpus/ comparing double precision of some of the NVIDIA systems

• GeForce GTX 1080 Ti up to 0.355 TFLOPS
• GeForce Titan Xp up to 0.380 TFLOPS

According to http://www.tomshardware.com/answers/id-3458667/nvidia-quadro-p4000-double-precission.html the P4000 double precision would be 0.166 TFLOPS

But for example the Quadro GP100 is 5.2 TFLOPS double precision, but lower single precision than the GTX 1080 Ti.

Thus it is important to know your computation mix. (And your budget. And your heat budget. And what kind of slots are needed and whether you will have to buy a complete new computer to use the device...)

[Joss Knight]

Yes, you're quite right, thanks Walter.

[Roberto]

Considering that I have no reason to use double precision for deep learning tasks and that for standard shallow nets of small-medium size the GPU is not much better than the CPU, I think at this point that 1080ti (or titan xp) is the only suitable option for the <1000$ price range. Thank you guys!