GPU benchmark for deep learning?

[martward]

Hi all,

 

I was wondering if anyone knew of a benchmark that one can use to compare GPUs for their performance on computations for things as Deep Learning (Cuda / OpenCL). I can't imagine that game performance scales well with performance on Deep Learning tasks.

 

Thanks

[minibois]

Maybe something like rendering in Adobe Premiere/Sony Vegas can be a sort of benchmark for it. Too bad Sony Vegas only supports AMD and Adobe Premiere only NVidia.. (at least the last time I checked it was that case)

[martward]

1 minute ago, Minibois said:

Maybe something like rendering in Adobe Premiere/Sony Vegas can be a sort of benchmark for it. Too bad Sony Vegas only supports AMD and Adobe Premiere only NVidia.. (at least the last time I checked it was that case)

Hmm yeah. Other problem is that these are both not open source, so that would be a pretty expensive benchmark for someone who doesn't use those programs otherwise. And actually writing a program that supports CUDA and OpenCL just to benchmark this, which most people don't care about anyway, seems like a lot of trouble. 

[Qwweb]

18 minutes ago, martward said:

I was wondering if anyone knew of a benchmark that one can use to compare GPUs for their performance on computations for things as Deep Learning (Cuda / OpenCL). I can't imagine that game performance scales well with performance on Deep Learning tasks.

Deep learning is a completely different animal, one that is highly reliant on single and especially dual precision and raw compute numbers (for number crunching). Dual precision FLOP performance numbers are usually a good indicator of DL performance, keep in mind that only the Tesla and Quadro cards have this capability (for dual precision) though. Deep Learning has very few applications unless you absolutely know what you want it for.

[martward]

Just now, Qwweb said:

Deep learning is a completely different animal, one that is highly reliant on single and especially dual precision and raw compute numbers (for number crunching). Dual precision FLOP performance numbers are usually a good indicator of DL performance, keep in mind that only the Tesla and Quadro cards have this capability (for dual precision) though. Deep Learning has very few applications unless you absolutely know what you want it for.

I have a 1060 and use it solely for TensorFlow atm, but a friend of mine wants a GPU for Deep Learning as well (TensorFlow and Theano) and he is doubting between the 1060 and the 1050TI, since the price gap is pretty big and he doesn't play games. 

I just found CompuBench, not sure if it's any good though.

[Qwweb]

3 minutes ago, martward said:

I have a 1060 and use it solely for TensorFlow atm, but a friend of mine wants a GPU for Deep Learning as well (TensorFlow and Theano) and he is doubting between the 1060 and the 1050TI, since the price gap is pretty big and he doesn't play games. 

I just found CompuBench, not sure if it's any good though.

What do you guys do with it? That will likely answer your question. I would suggest looking into Nvidia's Tegra-based developer boards (~$250-$500) which were specifically designed for deep and machine learning development. Nvidia has a guide for developers in regards to GPUs, integrated systems, and cloud-based services built for deep learning.

[martward]

16 minutes ago, Qwweb said:

What do you guys do with it? That will likely answer your question. I would suggest looking into Nvidia's Tegra-based developer boards (~$250-$500) which were specifically designed for deep and machine learning development. Nvidia has a guide for developers in regards to GPUs, integrated systems, and cloud-based services built for deep learning.

I'm doing a thesis in Information Retrieval and het is doing one with Computer Vision. Thanks, I will have a look at those boards :).

[Qwweb]

Just now, martward said:

Computer Vision.

I was on a robotics team that used the Jetson TK1 for robot vision co-processing, it is a great board for the price. Cloud based services are also a great resource if you need extra horsepower due to the cost effectiveness and usage of a tesla accelerated server.

[martward]

Just now, Qwweb said:

I was on a robotics team that used the Jetson TK1 for robot vision co-processing, it is a great board for the price. Cloud based services are also a great resource if you need extra horsepower due to the cost effectiveness and usage of a tesla accelerated server.

Not a huge fan of Cloud based services when I am still writing the code, it makes debugging and testing a whole lot harder than on a local PC. Also if you have a lot of writes to a file then it can become quite slow on certain types of servers (at least the ones we had access to) making my 1060 system outperform their Tesla system :P.