LinusTechTips should do deep learning benchmarks

[ratzes]

Just saw the compensator build and how the benchmarks were all wasted on gaming. Deep learning to my knowledge is one of the few fields where 4x or 8x titans actually makes a really big difference. It would be awesome to pull some useful knowledge about what a 4x titan build can scale to over a 1x or 2x build. Deep learning is also one of the few cases where I've seen a legitimate use for water cooling as the thermal throttling directly affects the time to train over the weeks they might run and isn't just an fps drop. It would be really cool to see some super useful data come out of these $10k systems as deep learning is almost entirely gpu bound for the right implementations with async mem copy. Can also show some linux/linus love

[Enderman]

I wish I could get 4 titan XPs to do all my deep learning for me rather than going to university.

[Guest]

12 minutes ago, Enderman said:

I wish I could get 4 titan XPs to do all my deep learning for me rather than going to university.

But you can buy all those cards with the uni tuition

[vanished]

To be honest, it just seems like something a little beyond and outside of what they typically cover.

[Enderman]

2 minutes ago, deXxterlab97 said:

But you can buy all those cards with the uni tuition

I could buy way more than that with my university tuition.....

[ratzes]

10 minutes ago, Ryan_Vickers said:

To be honest, it just seems like something a little beyond and outside of what they typically cover.

Yeah, I could see that, but we still get some nice cpu scaling feedback from a prime finding program.

[fjdhdhcyfhf]

how about folding? https://en.wikipedia.org/wiki/Folding@home#Graphics_processing_units

[vanished]

8 minutes ago, SCHISCHKA said:

how about folding? https://en.wikipedia.org/wiki/Folding@home#Graphics_processing_units

that would be glorious... a single Titan XP can average about 1.4 million PPD and because each card runs independently, scaling is 100% :drools:

[ratzes]

One more argument for deep learning/neural net benchmarks is that a decent amount of people are actually buying the cards to be able to do this (Nvidia marketed the titan xp for this as well). People are doing research, along with running stuff like neural style for images (https://github.com/jcjohnson/neural-style), where a 1600x1200 image can take 20 minutes to run using 10 gigs of vram.

[vanished]

26 minutes ago, ratzes said:

One more argument for deep learning/neural net benchmarks is that a decent amount of people are actually buying the cards to be able to do this (Nvidia marketed the titan xp for this as well). People are doing research, along with running stuff like neural style for images (https://github.com/jcjohnson/neural-style), where a 1600x1200 image can take 20 minutes to run using 10 gigs of vram.

oh wow is that cool.  I want to see it used on music now, where a midi file takes the place of the photograph and a song in a style you like takes the place of the art  

[Qwweb]

17 hours ago, ratzes said:

Just saw the compensator build and how the benchmarks were all wasted on gaming. Deep learning to my knowledge is one of the few fields where 4x or 8x titans actually makes a really big difference. It would be awesome to pull some useful knowledge about what a 4x titan build can scale to over a 1x or 2x build. Deep learning is also one of the few cases where I've seen a legitimate use for water cooling as the thermal throttling directly affects the time to train over the weeks they might run and isn't just an fps drop. It would be really cool to see some super useful data come out of these $10k systems as deep learning is almost entirely gpu bound for the right implementations with async mem copy. Can also show some linux/linus love

Deep learning typically ditches Titans (which are typically used for training) and goes for Tesla/Tegra (on the Nvidia side) in server based accelerator arrays. The more FLOPS the better and Tesla arrays now range well into the petaFLOP range. I'm not saying that there wouldn't be results, but stacked against the results of something throwing a few petaFLOPS around, teraFLOPS are gonna look insignificant. Tesla GPUs also work in parallel allowing for up to 16 cards totaling a massive 61440 CUDA cores running in parallel opposed to the Titan's 14336 (unstable using SLI).