Running Some New Benchmarks On The GTX 1080 (Octane Render, OpenCL)

[sdack]

I was about to ignore the article, seeing how a 780Ti beats a 980 and a Titan X in one of the tests.

Should I care?! ...

[edwaleni]

Things are looking really good for GTX 1080. I am sure a subset of Phoronix readers would be interested in 1080's deep learning performance. There are indications that 1080 does not have double-rate FP16 performance (compared to FP32) but rather 1/64, so we'll be left with whatever advances were made in FP32 speed vs. Maxwell. Are you planning to test AlexNet anytime soon and post the results? A comparison to 980, 980Ti and Titan would be great.

That subset happens to include me.

[Michael]

Things are looking really good for GTX 1080. I am sure a subset of Phoronix readers would be interested in 1080's deep learning performance. There are indications that 1080 does not have double-rate FP16 performance (compared to FP32) but rather 1/64, so we'll be left with whatever advances were made in FP32 speed vs. Maxwell. Are you planning to test AlexNet anytime soon and post the results? A comparison to 980, 980Ti and Titan would be great.

That subset happens to include me.

I will be putting out some of those numbers soon if my Caffe AlexNet test profile still is working.

If anyone has any other deep learning benchmarks that are easy to setup and can be automated/scripted and don'r require an ugly mess of dependencies, happy to incorporate them.

[torsionbar28]

Michael, do you still have a Titan Black (Kepler)? This card is not crippled for FP64 and I'd love to see the comparison with 1080. I'm surprised to see better performances for the 780 compared to 980 in OpenCL FP64.

Thanks for your work!

x2, I think the Titan and Titan Black are the only non-Quadro cards from Nvidia that aren't crippled for FP64. They would blow the 1080 out of the water, for sure, but it would be nice to see by how much. I'm running a GTX Titan specifically for its FP64 performance, as it's got Quadro level FP64, but without the Quadro price tag.

[vocoloco]

I will be putting out some of those numbers soon if my Caffe AlexNet test profile still is working.

If anyone has any other deep learning benchmarks that are easy to setup and can be automated/scripted and don'r require an ugly mess of dependencies, happy to incorporate them.

Thank you for doing the benchmark! The results are very exciting, even sans double-rate FP16. Regarding other frameworks, that would be certainly very interesting. There are a couple Github projects out there that you could probably leverage without much effort: https://github.com/soumith/convnet-benchmarks and https://github.com/glample/rnn-benchmarks.