Machine Learning: Titan Xp vs 1080ti

Zyndo · May 20, 2017

Title pretty much says it all. A guy i know (well not really, its my brother's friend's client... yeah) is looking to build a Machine learning rig. No idea of the budget or the criteria involved, but I have been asked for my advice. I don't really know anything about Machine Learning specifically. I mean, I know what it is, but I do not know exactly what it entails as far as hardware requirements goes. I've been asked for a way to improve the build or reduce the cost, but without knowing exactly how the hardware will be used or leveraged that is somewhat difficult. I've made alterations and recommendations for other parts (such as motherboards, SSD's, and so on). I know that a 1080ti is very close to a Titan Xp as far as raw compute power goes (12 TFLOPS on the titan vs 11.3 TFLOPS on the 1080ti), and this could be a good way to save some cash, but what I don't know is if the Titan Xp gives some unique advantage which makes it irreplaceable against the 1080ti, despite its massive price tag.The guy wants to put 4 Titan Xp's in the clients rig, so putting 4 1080ti's in there instead could save him thousands (CAD) and cost him very little in terms of overall performance.

I know this may be a difficult question to answer under those circumstances but I shall ask anyway:

Is a 1080ti is fully capable of doing anything a Titan Xp can do, just at a slightly lower level of performance?

All of my research and understanding of the Titan Xp and 1080ti shows that there is nothing truly unique between them. The Titan Xp is merely more powerful than a 1080ti, and comes with 1 additional GB of VRAM. Assuming there is no need for that 1 additional GB of VRAM, is there any other reason a 1080ti would be unable to do anything a Titan Xp could? Its my understanding that with some hardware and the way its designed, gives it significant advantages over its competitors even beyond its raw TFLOPS of performance or computational abilities. A good example of this would be ECC memory. If you NEED your ECC memory for some specific reason or task, then any amount of normal RAM (no matter how much faster or slower it is, or how much capacity you have) simply will not do.

I seek to know if the Titan Xp has any unique features about it, which gives it an irreplaceable advantage over the 1080ti (in reference to "machine learning and complex simulations"). Everything I know about these two cards would lead me to believe not, but I also certainly don't know everything, and almost nothing about machine learning.

I'm not looking for "oh this is a waste of money" or "just wait 'x' amount of time and you can get brand new 'y' card" or any irrelevant information like that. I'm merely looking for any unique distinctions between these two in the field of machine learning. any information along that long would be very much appreciated.

maybethisnamewillwork · May 20, 2017

1 hour ago, Zyndo said:

Is a 1080ti is fully capable of doing anything a Titan Xp can do, just at a slightly lower level of performance?

Yes, Nvidia has gimped the FP64-performance since the original Titan (GK110), so there's nothing "special". All you really pay is the early adopter tax.

Zyndo · May 20, 2017

Just now, maybethisnamewillwork said:

Yes, Nvidia has gimped the FP16-performance since the original Titan (GK100), so there's nothing "special". All you really pay is the "early adopter tax".

Okay this confirms what I heard, although I do not know the specifics of what FP-16 is. Any brief rundown you could make of the FP-16 rating? or how it relates to GPU performance? Or maybe direct me to a link where I could read up on that myself? That would be greatly appreciated

Zyndo · May 20, 2017

8 minutes ago, GamingMemeKing said:

I posted a similar question a few weeks ago.

From what everyone said in the posts, besides the VRAM, it was just a marketing ploy really.

and by VRAM you mean the 11GB capacity on the 1080ti vs the 12GB on the titan? Yeah that could be a deal breaker, but its such a small amount of difference that I doubt it would matter. 11GB is already a crapton... and from my very limited understanding of machine learning and compute tasks, VRAM really isn't all that important (especially since this isn't SLI, so its not like its 11GB shared between the 4 cards, but probably rather 44GB vs 48GB depending on how he sets up his programs and whatnot)

Zyndo · May 20, 2017

50 minutes ago, GamingMemeKing said:

Ahhh, well... From what i remember, multiple cards doesn't mean additional VRAM.

So whether it is 1 card, or 4 cards, you will still only have 11 or 12 GB of VRAM.

(However, 11GB is quite a lot.)

well that is how it works for crossfire/SLI... those specific technologies result in shared VRAM, but that doesn't mean that ALL ways to utilize multi GPU will result in shared VRAM. But I also really don't know that this would be one of those times. But either way, you're right, 11GB is still plenty regardless

maybethisnamewillwork · May 20, 2017

1 hour ago, Zyndo said:

Okay this confirms what I heard, although I do not know the specifics of what FP-16 is. Any brief rundown you could make of the FP-16 rating? or how it relates to GPU performance? Or maybe direct me to a link where I could read up on that myself? That would be greatly appreciated

Oops, I was referring to FP64 (double-precision).

Basically single precision floating point deals with 32-bit floating point numbers and double precision deals with 64-bit.

The number of bits in double precision increases the maximum value that can be stored as well as increasing the precision (the number of significant digits).

Here's a good example about usage of single precision vs double precision:
http://www.tuflow.com/forum/index.php?/topic/821-single-precision-vs-double-precision/#comment-2090