[Kitguru]Nvidia`s big Pascal GP100 have taped out - Q1 2016 release

[Cloudfire777]

Kituru writes:

Nvidia Corp. has reportedly taped out its next-generation high-performance graphics processing unit that belongs to the “Pascal” family, according to a market rumour. If the information is correct, then Nvidia is on-track to release its new GPU around mid-2016. The company needs its “Big Pascal” graphics processor to build next-generation Tesla accelerators for high-performance computing applications and better compete against AMD on the market of consumer GPUs

An anonymous person presumably with access to confidential information in the semiconductor industry revealed in a post over at Beyond3D forums that Nvidia had already taped out its next-generation code-named GP100 graphics processing unit. Nowadays, a tape-out means that the design of an integrated circuit has been finalized, but the first actual chips materialize only months after their tape-out.

Beyond3D

Time to revive this thread. My info says big Pascal has taped out, and is on TSMC 16nm (Unknown if this is FF or FF+, though I suspect it is FF+). Target release date is Q1'16. This is a change from Kepler and Maxwell where the smallest chip (GK107 and GM107 respectively) taped out first. Maybe the experience with 20nm was enough for NV to go back to their usual big die first strategy. Given the huge gains in performance compared to 28nm, and the fact that the 16nm process is both immature and quite expensive, I suspect the die size may be a bit smaller than what we've seen with GK110/GM200.

Simulatinously, there was a thread made on Chiphell today that said GP100 will be first on Pascal with a Tesla card featuring this with 8/16GB HBM2 memory.

Here is a top view of a Pascal card (prototype). Remember that with HBM2 those stacks are 2GB instead of 1GB (HBM1).

 

[Cloudfire777]

Not impossible this is one of the reasons why Nvidia launched GTX 980Ti so quick after Titan X. They got a new beast coming out already in February next year based on Pascal

 

[as9hW]

That package does not seem to have an interposer, or am I missing it?

 

[Mondozei]

I'd be interesting to see if we will even see GP100 in 2016 at all (for consumers)

I mean, the amount of new stuff in Pascal is staggering. The GP104 could very well mean a 50+% improvement over GM200 on its own.
So why rush GP100 in such a scenario? 2017 GP100, 2018 new node and/or new arch.

 

[Tuna-Fish]

Not impossible this is one of the reasons why Nvidia launched GTX 980Ti so quick after Titan X. They got a new beast coming out already in February next year based on Pascal

Taped out today on a new process doesn't mean release in February. If this info is correct, expect Q2.

 

[Tuna-Fish]

I'd be interesting to see if we will even see GP100 in 2016 at all.

I mean, the amount of new stuff in Pascal is staggering. The GP104 could very well mean a 50+% improvement over GM200 on its own.
So why rush GP100 in such a scenario? 2017 GP100, 2018 new node and/or new arch.

Currently they lack a top end card for the professional market with good 64bit perf. Keplers don't have the power efficiency gains of Maxwell, but all Maxwell chips have anemic 64-bit units.

I expect that GP100 will be initially sold at huge markups as Quadros and Teslas, with only a few sold to gamers as the Titan XX or whatever.

 

[Deders]

Has it been confirmed that pascal will be a gaming gpu or will it just be for datacentres?

I know it states it is dx12 compatible but then so are quadro's.

 

[ocre]

I'd be interesting to see if we will even see GP100 in 2016 at all (for consumers)

I mean, the amount of new stuff in Pascal is staggering. The GP104 could very well mean a 50+% improvement over GM200 on its own.
So why rush GP100 in such a scenario? 2017 GP100, 2018 new node and/or new arch.

This

Nvidia will need to release a card with DP. I am sure they will try to get it out as soon as possible. But these will be for tesla and quadro. Very expensive.

Big pascal for consumers/gamers won't be reasonable for a long time. I would say q4 2016 or later even.

 

[Silverforce11]

I haven't read much public info regarding TSMC's 16nm, but since they sucked for the initial 40nm & 28nm, blew donkey on 20nm.. we're somehow supposed to believe they can manage BIG Pascal, with a new uarch, new ram, stacking all in one on an unproven node... no. Tell 'em they dreaming.

Almost everyone who has paid attention will know NV needs to refresh their Tesla stack due to Maxwell's neutered DP compute, so the first big Pascal going to Tesla is a no brainer.

 

[Erenhardt]

That package does not seem to have an interposer, or am I missing it?

Yea. and seems like it is rendered. The HBM is a piss poor paint copy/paste.

 

[DarkKnightDude]

We probably won't see it till spring.

 

[nvgpu]

http://www.tsmc.com/tsmcdotcom/PRListingNewsAction.do?action=detail&newsid=9001&language=E

November 12, 2014

TSMC (TWSE: 2330, NYSE: TSM) today announced its 16-nanometer FinFET Plus (16FF+) process is now in risk production. This enhanced version of TSMC’s 16FF process operates 40% faster than the company’s planar 20-nanometer system-on-chip (20SoC) process, or consumes 50% less power at the same speed. It offers customers a new level of performance and power optimization targeted at the next generation of high-end mobile, computing, networking, and consumer applications.

The 16FF+ process is on track to pass full reliability qualification later in November, and nearly 60 customer designs are currently scheduled to tape out by the end of 2015. Due to rapid progress in yield and performance, TSMC anticipates 16FF+ volume ramp will begin around July in 2015.

https://www.youtube.com/watch?v=wFSxj_Msc70

4:20, Jen-Hsun holds the Pascal test module.

 

[tviceman]

So much of this rumor is unbelievable. For it to be real, there are three major and unlikely IF's that need to line up.

1) Finfets are getting better than expected yields at TSMC
2) Pascal GP100 isn't as big as either GK110 or GM200
3) Nvidia isn't leading a new family of GPU's with the small die first

GK107 (GK107 was released in notebooks a few weeks prior to GK104) and GM107 were both released first among their respective family of chips. GM107, while still without an equal in the < 75 watt segment, is starting to get long in the tooth in performance. That, coupled with the fact that Nvidia hasn't released a flagship die smaller 500mm2 since GT200b (2007), coupled with TSMC's track record, makes this rumor very hard to believe.

I have no doubt that if Nvidia goes with finfet +, and Pascal in itself is more efficient than Maxwell, we'll be looking at 2x performance (or slightly more) for GP104 over GM204. I have no idea how GP100 will fare over GM200 since GM200 has no transistors dedicated to double precision and GP100 likely will.

 

[Grooveriding]

We probably won't see it till spring.

Given TSMC's record for the past two nodes of being late I wouldn't be surprised at a Winter 2016 launch.

This is an interesting time with GPUs and how long we've been stuck getting products from the same 28nm node. Both companies have had to try and get more out of the same node. Nvidia managed to get more gaming performance in Maxwell by gutting it of the usual pro features like DP. Now they're going to have to keep using Kepler for their DP prodcuts until they can launch Pascal mid to late 2016.

Interested to see how AMD's Fury performs and if some of the rumours of it being faster than GM200 are true. Also if it raises the bar in DP performance. AMD might gain some share in the pro market if it does because nvidia has nothing but Kepler. Or AMD may have gone the same route as Maxwell and just focused on a pure gaming chip.

Usually it's AMD who has the performance crown at the start of a new node and then nvidia takes it back and holds it until the next node. We may see AMD with the performance crown for at least a year if Fury delivers as nvidia only has GM200, making Titan X their max performance, until 16nm can produce enough chips to bring Pascal to market. Also if Apple and some of the other larger companies want in on TSMC's 16nm, they are going to be taking whatever is available initially.

 

[gamervivek]

So that explains why it was getting updated so much while AMD thread had died down despite the latter being close to release.

 

[ShintaiDK]

Yea. and seems like it is rendered.

Lifelike rendering then.

 

[el etro]

GP204 obviously comes first, and when the yields get better, GP200 comes. The process simply had too many years to get mature, is about time the 16FF+ gets done.

AMD and Nvidia didn't shrunk to any newer node these times because there don't have any newer high performance node to shrunk. They were simply using the best node they can use.

I really don't know if GF14LPP can be used to high performance devices, but TSMC16FF+ surely can. Link: http://community.cadence.com/cadenc...instream-designers-and-internet-of-things-iot

 

[tviceman]

GP204 obviously comes first

Why obviously? GK107 and GM107 both came before their GXX04 counterparts.

 

[96Firebird]

Lifelike rendering then.

He must be holding a hologram of the rendering... Which means Nvidia is way ahead of the curve on technology!

 

[el etro]

Why obviously? GK107 and GM107 both came before their GXX04 counterparts.

First than GP200, since Nvidia is waiting better Yields to make it happen. obviously GP107 can come first than GP204...

 

[jpiniero]

First than GP200, since Nvidia is waiting better Yields to make it happen. obviously GP107 can come first than GP204...

The idea is that it wouldn't be a "Big" die. Like 20% faster than Titan X but with 1/2 DP.

 

[el etro]

The idea is that it wouldn't be a "Big" die. Like 20% faster than Titan X but with 1/2 DP.

Or twice the TTX with the DP enabled.

 

[maddie]

The timing of this is laughable.

Has anyone released any product on TSMC Finfet+ process as yet?
Is HBM2 ready?

We are still expected to believe BIG Pascal in Q1 2016?

I thought the supporters of AMD were supposed to be delusional.

 

[MarkLuvsCS]

I doubt initial dies are going to go into any consumer products. There's an obvious reason they are able to invest money even if lower yields doesn't make sense for AMD because of the supercomputer projects they are involved in:

http://www.anandtech.com/show/8727/nvidia-ibm-supercomputers

Remember the original Titan was released for consumers about 5-6 months after the Titan supercomputer received their Telsas which utilized the same GPU Core. So expect BIG Pascal Late Q2, early Q3.

 

[Cloudfire777]

Taped out today on a new process doesn't mean release in February. If this info is correct, expect Q2.

Try reading the article again. Even the headline should be enough.

They are aiming for Q1 release.

16nm FinFET have been available for Nvidia for a long time, to make prototype designs etc, just not manufactured in large scale. 16nm process willbe available for mass production for Nvidia and the rest of TSMCs clients very soon if TSMC havent already begun.
Edit: nvgpu posted a link that says volume production in July this year. Next month

With AMD very soon out with HBM, you better be damn sure Nvidia is eager to get Pascal out with HBM too as soon as possible.
Maxwell was never designed for 28nm anyway. It was meant for 20nm