Nvidia Tesla V100 with Volta GV100 GPU announced

[HEXUS]

The GPU packs in 21.1 billion transistors and is fabricated using TSMCs 12nm FFN process.

Read more.

[UseItNow]

$149000?

I wonder what a used Tesla K40 fetches these days?

[qasdfdsaq]

$149000?

I wonder what a used Tesla K40 fetches these days?

Quite a lot, given the Tesla K40 is a generic PCIe add-in card you can bung into any modern workstation or server, whereas the $149,000 DGX-1 is a ultra-high-end proprietary supercomputer module with eight Tesla V100's in it

[watercooled]

Wow that die is simply monstrous. I don't even recall Intel making something that large! I assume they're talking about the logic die rather than the interposer, but I thought TSMC's single-exposure reticle size was somewhere around the 600mm2 mark, hence both AMD and Nvidia hitting that wall on 28nm. IIRC they can get around that limit for things like interposers by using multiple exposures but that wouldn't work for a logic die (I could imagine a die somehow divided down the middle working, but then it would make no sense to leave it as a singe die vs MCM).

I wonder what the real motivation is behind something like this? It seems like it's more marketing and halo effect than anything - they could have gone marginally smaller with a relatively insubstantial difference in performance but massive improvements to costs and yield. And for the sort of application this seems to be targeted at, even Nvidia are showing it must scale well across nodes given they're sticking eight of these in a box. Are they just going for no-expense-spared bragging rights against the Xeon Phi for die size and performance? Although it must be noted Intel are being very secretive around Phi's die size, which some have estimated to be in the area of 700mm2 IIRC. Then again it's not targeting exactly the same market as Phi has access to far more memory than this. This seems like something with a very niche market, but I guess there must be some demand for it to pour this much money into producing it!

[DirtyDean68]

Am I right in thinking that Volta GeForce GPU cards can't be far off? 3 months away at a guess.

[UseItNow]

$149000?

I wonder what a used Tesla K40 fetches these days?

Quite a lot, given the Tesla K40 is a generic PCIe add-in card you can bung into any modern workstation or server, whereas the $149,000 DGX-1 is a ultra-high-end proprietary supercomputer module with eight Tesla V100's in it

Haha, probably should have worded that better...

Did Google the DGX-1, just wondering what they'll price those V100's at?

[mtyson]

Tesla V100 announcement presentation segment

[kalniel]

Am I right in thinking that Volta GeForce GPU cards can't be far off? 3 months away at a guess.

We'll have to see! Pascal GPUs were quite different from the Tesla P100 though.

[kompukare]

Am I right in thinking that Volta GeForce GPU cards can't be far off? 3 months away at a guess.

Very unlikely. H1 2018 is more likely with maybe March 2018 being the most likely. No idea what the yields and potential volume of TSMC's 12nm process is but at $14,000 / 8 is still $1,750 so bad yields are unlikely to deter them for GV100 but would do so for GV104 etc. Actually, $1,750 isn't that bad for 800mm2 going by Nvidia's usual margins so they must be desperate to keep their HPC customers against competition from (for example) Intel's Phi.

[azrael-]

Did anyone spot any woodscrews?

[The Khemist]

Can it run Crysis that is the real question? ...... sorry I will get me coat!

[chj]

Wow that joke is 10 years old now

[watercooled]

We'll have to see! Pascal GPUs were quite different from the Tesla P100 though.

Yeah, the large dies aimed at Tesla products have become very different beasts than their gaming counterparts now. For example HBM vs GDDR memory controllers, presence of NVlink, and so on. It would appear that Nvidia's enterprise business is now big enough for them to make dedicated products for it; the current big dies are made specifically for this sector, and the largest gaming dies lack the sort of features that justified the original Titans.

As a comparison, the GK110 was quite big for its gaming performance as it was a sort of hybrid product with a lot of die space which was effectively unused in gaming. On the other end of the scale, the GM200 was only slightly bigger than the GK110, but offered substantially better gaming performance on the same node. The maturity of the node likely went towards making such a large dedicated gaming GPU possible.

[The Khemist]

Wow that joke is 10 years old now

Crysis 3 - more up to date version of the joke

[Xlucine]

Very unlikely. H1 2018 is more likely with maybe March 2018 being the most likely. No idea what the yields and potential volume of TSMC's 12nm process is but at $14,000 / 8 is still $1,750 so bad yields are unlikely to deter them for GV100 but would do so for GV104 etc. Actually, $1,750 isn't that bad for 800mm2 going by Nvidia's usual margins so they must be desperate to keep their HPC customers against competition from (for example) Intel's Phi.

I think you missed a character there - the DGX-1 system is $149,000 - this works out to $18,625 per GPU if we ignore the rest of it. Almost $19k for a 800mm^2 GPU is much more reasonable, if nvidia could sell 800mm^2 and make a profit at $1,750 then I'd be expecting an equal sized die in a titan product, like the titan Z of a few years ago