Titan Z vs K40 processor?

Question

I'm using GPUs for scientific computing. Recently Nvidia released its flagship product GeForce Titan Z. I would like to know, how this processor fairs against Tesla K40 (another NVIDIA product). I have already checked the specs but keen to know of any benchmarks between these two processors, or on the ability of Titan Z for scientific computing applications. I also would like to know if the Titan Z should be treated as single GPU or two GPU from the programming perspective.

Thanks in Advance, Regards, Sakthi K

Answer 1

Its got two chips; undoubtly it will act as two separate cards, from a compute perspective, like all other cards of this kind before it.

I have worked with titans and other NVidia gaming cards for scientific computing extensively over the last years, and they work just fine for my purposes, but as always, 'it depends'. First of all, if you absolutely do need double precision, then they are a bad deal. Of course most applictions, including scientific simulations, are not actually constrained by the limits of single precision floats; but for some applications it does matter.

So the K40 has more memory per chip, and more double precision performance. But if you are sure you dont need either of those (like I do for my next build), a pair of Titan Z's is a pretty good way to cram an insane amount of single precision performance into a manageable form factor.

(edit: I see titan z unlike previous gaming cards has full double precision too; so if you do need double precision, that adds to its value. personally, I find memory more often limiting than fp precision though)

Answer 2

Titan Z is in essence two Titan Black chips on same card with individual 6GB VRAM. Main advantage of Titan Z is that two chips are synchronized so load will be distributed evenly and also that it costs ~$2100 vs cost of $1700 for single Titan Black. The Titan Black itself is designed for mainly single precision units, limited double precision units and relatively low end memory.

I would suggest Titan X is probably better choice than either Titan Z or Titan Black due to two reasons: (1) it costs only $1100 (2) it has twice the memory. So you can actually build much powerful system using two Titan X instead of one Titan Z with approximately the same cost. This is especially true for deep learning related work. In fact, NVidia Digits Devbox features Titan X as well.

If your computation requires double precision (FP64) then you need to take a look at Tesla chips like K40 which are almost twice as expensive and with same memory as Titan X. Most deep learning and machine learning related work involves single precision so K40 or K80 is not suitable compared Titan X.