Deep Learning & CUDA Benchmarks On The GeForce GTX 1080 Under Linux

[phoronix]

Phoronix: Deep Learning & CUDA Benchmarks On The GeForce GTX 1080 Under Linux

Last week I published the first Linux review of the GeForce GTX 1080 followed by some performance-per-Watt and OpenGL results from the GTX 1080 going as far back as the 9800GTX, among other interesting follow-up tests with OpenGL/Vulkan/OpenCL. Since then one of the most popular requests has been for doing some deep learning benchmarks on the GTX 1080 along with some CUDA benchmarks, for those not relying upon OpenCL for open GPGPU computing. Here are some raw performance numbers as well as performance-per-Watt in the CUDA space.

http://www.phoronix.com/vr.php?view=23281

[mastermind]

It would be interesting to see how it performs against Titan Z on FP64 (double precision math). Titan X (and 9xx) have very poor FP64 unit and that's what matters for machine learning. (unlike gaming graphics which is fine with FP32 / single precision).
See:
1. https://www.quora.com/Nvidia-GeForce...n-applications
2. http://www.anandtech.com/show/9059/t...tan-x-review/2

[Michael]

It would be interesting to see how it performs against Titan Z on FP64 (double precision math). Titan X (and 9xx) have very poor FP64 unit and that's what matters for machine learning. (unlike gaming graphics which is fine with FP32 / single precision).
See:
1. https://www.quora.com/Nvidia-GeForce...n-applications
2. http://www.anandtech.com/show/9059/t...tan-x-review/2

Unfortunately I don't have Titan Z....

[Michael]

Is there any benchmarks for OpenCL vs. CUDA on NVIDIA hardware?

You could compare my SHOC OpenCL results from my earlier GTX 1080 article to this one that's using CUDA version.

[chuckula]

It would be interesting to see how it performs against Titan Z on FP64 (double precision math). Titan X (and 9xx) have very poor FP64 unit and that's what matters for machine learning. (unlike gaming graphics which is fine with FP32 / single precision).
See:

Wrong. Many many machine learning applications actually require lower precision than even 32 bit single precison, which is why Nvidia went out of its way to double the 16-bit floating point performance in Pascal and why Google has been testing chips with even lower precision levels to enable higher performance using fewer transistors. Convolutional networks in particular include vast arrays of nodes that each receive a numeric weight value and in many instances updating the numeric weight values can be done with very low precision math while remaining useful.

You are thinking of other HPC aplications like CFD that actually require high precision math for most applictions. However, machine learning generally does not require high precison.

[dweigert]

Well, now I know that I need one of these cards, instead of a new bicycle...

[davidbepo]

Well, now I know that I need one of these cards, instead of a new bicycle...

better wait for rx 480

[starshipeleven]

Well, now I know that I need one of these cards, instead of a new bicycle...

I can buy around 5 bycicles with a 1080's worth of cash. And not the cheap crappy ones. Did Apple start making bikes while I wasn't watching?

[dweigert]

The cheapest bicycle I'm considering is $1200, on sale.... The most expensive bike in the shop was over $5K

[starshipeleven]

The cheapest bicycle I'm considering is $1200, on sale.... The most expensive bike in the shop was over $5K

Wow, 1000$ to over 4800$ of unjustified expense right there.

Unless it's a specialized device for competitive use anyway.