Pascal Tesla availability clue

[Arachnotronic]

From Cray's earnings call

On the development front, our roadmap for 2016 includes several exciting new upgrades to our existing products. We have three major supercomputing product releases currently planned for the year. Two based on new processors expected from Intel, the next-generation of the Xeon processor and follow-on to the Xeon Phi known as Knights Landing and a third, which will integrate the new Nvidia GPU, code name Pascal. A significant portion of our revenue this year will rely on these new processors as well as a number of other technologies. Much of it is expected to be recognized in the fourth quarter and the availability of these new processors are a key driver to our backend weighting for the year .

Pascal Tesla appears to be roughly TTM with Knights Landing. If Cray is shipping systems & recognizing significant revenue in 4Q, then GPU availability will need to be quite a bit before that.

 

[raghu78]

Lets wait till GTC 2016 and see if Nvidia can actually give a commitment on Pascal GP100 Tesla launch date. Right now its all guess work.

 

[Arachnotronic]

Lets wait till GTC 2016 and see if Nvidia can actually give a commitment on Pascal GP100 Tesla launch date. Right now its all guess work.

Seems like they've already given Cray, a major supercomputer vendor, a "commitment."

Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?

 

[Stuka87]

Seems like they've already given Cray, a major supercomputer vendor, a "commitment."

Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?

Tesla typically comes quite a while before a GeForce card. Sometimes over 6 months later.

 

[Arachnotronic]

Tesla typically comes quite a while before a GeForce card. Sometimes over 6 months later.

Well, in any case, it flies in the face of the rhetoric that some of the NVIDIA haters keep throwing around...

Significant revenue for Cray in Q4 means cards going to Cray at least in Q3, probably earlier given how complex these systems are to build. Anybody who thinks that NVIDIA hasn't taped out GP100 at this point is fooling his/herself.

 

[AtenRa]

Significant revenue for Cray in Q4 means cards going to Cray at least in Q3, probably earlier given how complex these systems are to build. Anybody who thinks that NVIDIA hasn't taped out GP100 at this point is fooling his/herself.

A significant portion of our revenue this year will rely on these new processors as well as a number of other technologies.

I Dont believe they are talking about the GPUs.

 

[The Alias]

Well, in any case, it flies in the face of the rhetoric that some of the NVIDIA haters keep throwing around...

Significant revenue for Cray in Q4 means cards going to Cray at least in Q3, probably earlier given how complex these systems are to build. Anybody who thinks that NVIDIA hasn't taped out GP100 at this point is fooling his/herself.

As a bonafide grade A lurker history would suggest the professional market gets cards way before the consumer market does. CRAY not getting significant profit from gp100 until q4 means that we probably won't get it until a long while after. Not to mention we don't know what percentage of that profit is coming from Knight's Landing as well. All that said, I sincerely hope gp100 and Vega come out as soon as possible as we all could use something substantial to talk about.

 

[Arachnotronic]

I Dont believe they are talking about the GPUs.

They mentioned three new processors: a new Xeon (Broadwell-EP), Xeon Phi mk2 (Knights Landing), and Pascal.

What processors do you think they were referring to if not the three above?

 

[lilltesaito]

They mentioned three new processors: a new Xeon (Broadwell-EP), Xeon Phi mk2 (Knights Landing), and Pascal.

What processors do you think they were referring to if not the three above?

From the quote:

We have three major supercomputing product releases

Two based on new processors

Third super-computing product is Nvidia GPU, code name Pascal.

 

[Arachnotronic]

From the quote:



Third super-computing product is Nvidia GPU, code name Pascal.

OK, but even so:

We have three major supercomputing product releases currently planned for the year.

Three releases for this year.

 

[AtenRa]

They mentioned three new processors: a new Xeon (Broadwell-EP), Xeon Phi mk2 (Knights Landing), and Pascal.

What processors do you think they were referring to if not the three above?

I said i dont believe they are referring to the GPUs (Pascal/Phi) when they are saying "A significant portion of our revenue" but more on the x86 XEONs.

 

[lilltesaito]

OK, but even so:



Three releases for this year.

I was just showing the break down for AtenRa comment.

But it also says, "rely on these new processors as well as a number of other technologies"

 

[ShintaiDK]

Good to see KNL and Pascal on track for the year

 

[xthetenth]

Seems like they've already given Cray, a major supercomputer vendor, a "commitment."

Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?

So if your reading is correct, Cray's going to be getting shipments and buying literally all they can through Q4. That rules out consumer GP100 coming out in 2016, doesn't it?

 

[MrTeal]

Anyone know if Cray would be using actual Tesla cards, or would they use a custom solution like the Kepler in Oak Ridge's Titan?

Edit: NM, answered my own question. The current Kepler based CS-Storm uses PCIe Tesla cards.

 

[NTMBK]

Seems like they've already given Cray, a major supercomputer vendor, a "commitment."

Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?

First Kepler Tesla cards were based on GK104, let's not jump to conclusions

EDIT: Good find though, definitely should be a new Titan card by the end of the year!

 

[Samwell]

There are already many signs of Big Pascal coming in half a year, like this one:

it will be installed in the CEA’s Very Large Computing Centre[2] in Bruyères-le-Châtel, France, mid-2016 to cover expanding industrial needs.

It will have 2304 Intel® Xeon® E5 processors (Broadwell) with a total of 32,256 cores at 2.4 Ghz and 18 hybrid nodes with Nvidia Pascal processors, for remote computing and visualisation.

http://www.bull.com/ccrt-boosts-industrial-innovation-petascale-supercomputer-bull

I'm sure we'll see the presentation at GTC.

 

[Sweepr]

Nice to see big Pascal on track for release later this year.

 

[Arachnotronic]

There are already many signs of Big Pascal coming in half a year, like this one:

http://www.bull.com/ccrt-boosts-industrial-innovation-petascale-supercomputer-bull

I'm sure we'll see the presentation at GTC.

Oh snap, good find.

 

[AtenRa]

First Kepler Tesla cards were based on GK104, let's not jump to conclusions

EDIT: Good find though, definitely should be a new Titan card by the end of the year!

Kepler Tesla was GK110 released on November 2012, 6-7 months after GK104.

What you remember was the Quadro line that used the same GK104 chips.

I dont see a GP100 in retail in 2017.

 

[NTMBK]

Kepler Tesla was GK110 released on November 2012, 6-7 months after GK104.

What you remember was the Quadro line that used the same GK104 chips.

I dont see a GP100 in retail in 2017.

Tesla K10 was first. 2x GK104.

 

[Kenmitch]

Well, in any case, it flies in the face of the rhetoric that some of the NVIDIA haters keep throwing around...

Significant revenue for Cray in Q4 means cards going to Cray at least in Q3, probably earlier given how complex these systems are to build. Anybody who thinks that NVIDIA hasn't taped out GP100 at this point is fooling his/herself.

There is a difference between NVidia haters and those who hate the NVidia fud that is pumped around here. Antagonizing them with the highlighted statement isn't really the best option.

I'm sure NVidia has already taped out Pascal by now. It's just a matter of time before some indication of when they will launch something useful for us mortals....As usual I'm expecting leaks, statements, a lot of fud, close to when AMD's launch is imminent.

 

[maddie]

I believe that large Pascal will have a GPU die of around 400mm^2. Nvidia in the SC15 talk [Path to exascale computing] by William Dally, indicated big Pascal as having a 30mm X 30mm size.

This made no sense at the time as we refer to the GPU as the processing unit on the interposer, and even if you ignore the yield problem, 900mm^2 is beyond the reticle limit.

Now that they are saying they have on-die memory, which is also confusing in another way, as this tech should not be ready for commercial use.

The solution to these statements, is that Nvidia is referring to the entire interposer area and associated components as the GPU.

From this 30mm X30mm area and accounting for the 4 stacks of HBM2, we end up with a traditional measured GPU die area of around 400mm^2

 

[AtenRa]

Tesla K10 was first. 2x GK104.

Yea you are right, forgot about K10.

 

[raghu78]

Tesla K10 was first. 2x GK104.

exactly. how people quickly forget the past.

http://www.anandtech.com/show/5840/...s-gk104-based-tesla-k10-gk110-based-tesla-k20

http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last

A Pascal GP104 Tesla is possible late this year. But GP100 is going to be difficult. Nvidia have not even shipped a 100 sq mm FINFET GPU and people want a 500 sq mm FINFET GPU. At the 28nm node Nvidia launched desktop GK104 aka GTX 680 in late March 2012, Tesla K10 (2 x GK104) in May 2012 and Tesla K20 (GK110) in Nov 2012. We won't see GK104 till probably Q4 2016 because its going to use GDDR5X most probably.

http://www.anandtech.com/show/9762/hisilicon-announces-kirin-950-huawei

The funny part is TSMC's largest high volume FINFET product is A9X at 147 sq mm and Samsung's largest high volume FINFET product is A9 at 96 sq mm. TSMC was hitting 80% yields on Kirin 950 back in Sep 2015.

Anyway I think few people are way too optimistic about a 500 sq mm FINFET GPU launching in 2016 when we are yet to see even a 300 sq mm one.