Nvidia Volta Speculation Thread

Yes, I work in a University/Lab that among other thing work on machine learning, and we have already ordered several cards.
For us, this is not ridiculous, I had personally expected it to be a "Quadro GV100" before GV100 was released as a Titan card. And for the performance this card will give when training networks, $3000 is really not that bad.

 

So next will be Titan Xv for $1,500?

 

I am guessing a "GTX" Titan V without tensor and with much higher clocks plus GDDR6.

On another note, the newly released Geforce driver has Titan V in it's support list , so it won't be long before someone test this card with actual games.

https://us.download.nvidia.com/Windows/388.59/388.59-win10-win8-win7-notebook-release-notes.pdf

 

Didn't Jensen Huang at one point this year mention Volta's manufacturing cost per unit was pretty high?

 

Duh, its 815mm^2 die.

 

And SLI support since these seem not to have it.

 

Poor Volta ...

 

They have NVLink - a bridge is only 599.

 

No, it's not functional:

https://www.anandtech.com/show/1213...dia-titan-v-video-card-gv100-for-3000-dollars

 

I wonder if there will be driver product segmentation as well, artificially gimping performance? Yes the card is $3,000 but the Tesla is $10,000 and there's not much difference between the 2. I'm guessing since Tesla is designed mostly for multi-GPU (nvlink) and/or rack systems for HPC, that market won't really be impacted by a full GV100 single card option at 30% of the price.

 

We are left wondering what the uplift in gaming would be, Titan V already has 25% more FP32 performance, and 20% more memory bandwidth than TitanXP, plus whatever architectural enhancement there maybe. And we know GV100 is substantially better than GP100.

For comparison's sake, Quadro GP100 was also 20%~40% faster than Quadro GP102 in several applications (courtesy of more caches and registers per FLOP), despite the GP100 being 20% slower on FP32 performance.

Can we guess how much faster would a Titan V be in games compared to Titan XP?

 

No NVLink. No ECC as well probably?

I think they chose their price well. I want one (but I don’t have a use for it. )

 

I would say there's probably not enough PC gaming whales willing to drop 6K on two cards to warrant it, but the die probably doesn't even have the necessary SLI bridge interfaces...

NV always do that thing where they disable AA'd lines for example and maybe some other stuff as well (which they've been doing since the original Geforce released back around the turn of the millennium), but then this card does cost as much as their regular Quadros, so maybe not this time. *shrug*

I guess we know soon enough... Delivery on the 12th allegedly, that's just four days unless there's a leak from somewhere.

 

That would be my thinking as well unless SLI and NVLink are somehow backwards compatible. As these should be the same V100 chips, SLI over NVLink wouldn't be worth the space.

Also possible SLI is being replaced by PCIE the way AMD implemented Crossfire. Makes more sense with explicit multi-adapter including multiple IHVs and added PCIe lanes in HEDT segments.

 

TIme will tell, anyway i doubt this card is intended to run for gaming ( nor even raytracing or FP32 cases ).. so does it really need SLI ? no.. In fact for 1000$ less you buy 2x TitanXP ( and end with 9TFlops more )

 

Titan V is squarely aimed at ML/DL crowd, it was even announced at NIPS.

 

They got one part of the Titan brand right, namely you know that is the best (and very expensive).

Still, for each and every incarnation you have to figure what's it actually good for

 

Ah, interesting. Just saw the physical connectors on the card. Nvidia did not bother to talk to many people about Titan V.

 

Yeah it will be done similar to the Quadro GP100, designed as a 2-way NVLink.

 

Hmm I thought it had the physical NVLink connectors though, same way as Quadro GP100 albeit version 2.
It would make sense to be active as this card is designed for universities/labs primarily, and other devs interested in various CUDA function testing before/outside of HPC.
You can also buy the cards as a pair from Nvidia *shrug*.

Maybe semantics that its use is limited to certain CUDA related applications-libraries/frameworks/etc.