** The Official Nvidia GeForce 'Pascal' Thread - for general gossip and discussions **

[Rroff]

It's how the industry's economics work. Chip binning and subsequent different SKUs and market segments. Re: yields, it couldn't not be disastrous at 600mm2 on a new node.

I think some people are being overly doom and gloom 20nm planar has been knocking around a bit now and even finfets its likely in a better place than the jump to 40nm, etc. even with the extra complications to overcome compared to 28nm and above. (And that is despite some early issues with producing dies >3xxmm2).

 

[GAC]

In a month's time, "Let's play wait until September for the actual cards."

The launch of this gen has been, and continues to be, painful to live through (for those of us who need a new card pronto! Need, yes. Mine is dying ).

christmas you say.

 

[wunkley]

Can some nice person please summarise what the current situation is now? More specifically...

1. Do we have any actual reliable info now about specs and performance of 'mid' pascal or 'top' pascal?

2. Do we have any actual reliable info now about release dates ?

Call me lazy but I really can't face wading through the pages of guff. Ta

 

[EPCCare]

Can some nice person please summarise what the current situation is now? More specifically...

1. Do we have any actual reliable info now about specs and performance of 'mid' pascal or 'top' pascal?

2. Do we have any actual reliable info now about release dates ?

Call me lazy but I really can't face wading through the pages of guff. Ta

1. No.

2. No.

Only conjecture and educated guesses.

 

[Rroff]

Can some nice person please summarise what the current situation is now? More specifically...

1. Do we have any actual reliable info now about specs and performance of 'mid' pascal or 'top' pascal?

2. Do we have any actual reliable info now about release dates ?

Call me lazy but I really can't face wading through the pages of guff. Ta

We have some information on the core for the P100 HPC targeted part and that is about it coupled with what most people already had assumed that 16GB HBM2 cards aren't likely to be arriving this side of Christmas.

From the nVidia developer page it can be assumed that the architecture is largely based on a tweaked Maxwell but that it is a little more than just Maxwell with DP added back in.

 

[JediFragger]

We have some information on the core for the P100 HPC targeted part and that is about it coupled with what most people already had assumed that 16GB HBM2 cards aren't likely to be arriving this side of Christmas.

From the nVidia developer page it can be assumed that the architecture is largely based on a tweaked Maxwell but that it is a little more than just Maxwell with DP added back in.

Agreed. Volta will be where the gravy is at

 

[Rroff]

Agreed. Volta will be where the gravy is at

Volta seems to be almost purely being built for Summit and/or other super computer type installations - I reckon it will be quite awhile until consumers get anything like it.

 

[TheF34RChannel]

Dude this is just tesla, we were discussing and there's no point in focusing all the proccess to fp64 untill you have decided to make this card just for computational resources that i think is the case.

For gaming we will see in may/june i hope the real gaming chip, this one for sure is just for tesla.

I know but if big Pascal has that number of cores I don't see the GeForce cards having double that, or higher at all for that matter.

 

[pmc25]

Can some nice person please summarise what the current situation is now? More specifically...

1. Do we have any actual reliable info now about specs and performance of 'mid' pascal or 'top' pascal?

2. Do we have any actual reliable info now about release dates ?

Call me lazy but I really can't face wading through the pages of guff. Ta

1. Mid. Nothing. Top, P100 is top. Will show up as Titan equivalent Q1 / Q2 '17. Questionable if it'll be competitive against Pascal mid range, let alone AMD Polaris or Vega.

2. No. NV claim they will start shipping P100 in June ... but IF they hit this, it will be a handful of test samples to biggest HPC contractees and IBM. Late Q4 for real product delivery. Q1 '17 for OEM Tesla parts. Titan thereafter. GP104 probably paper launch at Computex and Q4 availability.

 

[Rroff]

I know but if big Pascal has that number of cores I don't see the GeForce cards having double that, or higher at all for that matter.

There are changes to the architecture (and attainable clock speeds) - it would be flawed to do a face value comparison of the number of cores on Pascal versus Maxwell then say well its not that much higher number so it won't perform that different.

It always tickles me that people think core numbers will keep scaling up indefinitely - with rumours of like double the cores when a design has already been scaled up significantly from its original incarnation - as you start to run into various pipeline inefficiencies if you just keep scaling the same design up to have more and more of everything.

 

[TheRealDeal]

There are changes to the architecture (and attainable clock speeds) - it would be flawed to do a face value comparison of the number of cores on Pascal versus Maxwell then say well its not that much higher number so it won't perform that different.

It always tickles me that people think core numbers will keep scaling up indefinitely - with rumours of like double the cores when a design has already been scaled up significantly from its original incarnation - as you start to run into various pipeline inefficiencies if you just keep scaling the same design up to have more and more of everything.

There's already articles with the thinking that it's just Maxwell shrunk with compute added in so maybe there is not much in the way of architecture improvements. What NV do have in there pocket is clock speeds. An overclocked gtx980ti can gain around 30% on a stock card. Add in the extra Shaders and whatever else they beef up you could have another 30%. I still think Amd are looking better all round but in pure speed terms Nvidia will be up there as usual.

 

[JediFragger]

Volta seems to be almost purely being built for Summit and/or other super computer type installations - I reckon it will be quite awhile until consumers get anything like it.

D'oh!!!!

Ah well, 980Ti SLi best last me a while then!!!

 

[pmc25]

http://www.bitsandchips.it/52-english-news/6816-rumor-tsmc-will-produce-the-huge-gp100-interposer

Interposer is 1200mm2 (Fiji is 1000mm2).

Whole thing is being packaged by TSMC (their first attempt at a large interposer), not UMC / Amkor as with AMD.

The whole thing looks like it's set up to be a fiasco.

 

[Stanners]

http://www.bitsandchips.it/52-english-news/6816-rumor-tsmc-will-produce-the-huge-gp100-interposer

Interposer is 1200mm2 (Fiji is 1000mm2).

Whole thing is being packaged by TSMC (their first attempt at a large interposer), not UMC / Amkor as with AMD.

The whole thing looks like it's set up to be a fiasco.

Rumor and no named source?

 

[drunkenmaster]

http://www.bitsandchips.it/52-english-news/6816-rumor-tsmc-will-produce-the-huge-gp100-interposer

Interposer is 1200mm2 (Fiji is 1000mm2).

Whole thing is being packaged by TSMC (their first attempt at a large interposer), not UMC / Amkor as with AMD.

The whole thing looks like it's set up to be a fiasco.

The part of the conference where he was saying that as engineers one of their main guiding ideals throughout the company was never release a product which requires three 'miracles', using miracles instead of the more accurate, 3 completely new technologies, to come together to deliver a new chip. It's tick tock of Intel essentially. You don't want to do a new process, a new architecture and say a new memory technology all on one process.

Then proudly stated, this one is relying on 5... even though we know what a ridiculously bad idea that is as we've done our best to avoid 3 because it spells trouble, look how brave we are, we're doing something truly retarded, round of applause for our stupidity please.

Another stupid part was the algorithms, he literally spent an hour before this banging on about how many uber complex algorithms deep learning was using all of which can run faster on a card with more deep learning performance... so that isn't a 'miracle' required at all. On the other hand he simply added CoWoS with HBM2 as a single miracle. Yeah, making a chip to work on interposers, making interposers, using HBM2, establishing a supply chain for all the above and getting it all working as a high yielding packaging operation... that all counts as a single miracle when it's like 58.

Now don't get me wrong, this is about eleventy billion times easier because AMD has been working with a dozen or so companies and going to industry panels/conferences to talk about that technology for the past decade. AMD absolutely pioneered this and as such TSMC and other companies have a huge amount of help and knowledge from what other companies have done to achieve it.

AMD sensibly did all of this on a know architecture with a known process, and after a decade of basically coming up with all the technologies alongside these companies in the first place.


It just boggles the mind that he went up on stage and laid out precisely how stupid they were being like it was a great thing and not a disaster waiting to happen.

 

[wunkley]

Thanks for the summary fellas. Glad I bought a second 980Ti to SLI last week now. Perhaps a bit silly, but I had a strong feeling we wouldn't be seeing anything new that's worth getting excited about anytime soon.

 

[Rroff]

1200mm2 is about the minimum size to fit 4 stacks of HBM2 and a 610mm2 core - not much in the way of margin for error though but tends to lend some credibility to the rumour.

 

[8 Pack]

1. No.

2. No.

Only conjecture and educated guesses.

And this is why the entire thread is almost pointless.... Those who actually know can't say.......and the rest simply don't know.........

 

[TwsT]

How are people talking about the expected yields manufacturers will achieve in production and what they will do with the parts that don't meet the grade with such authority?

Is this genuinely people in the know or is it just conjecture, is this people who think they know what they're talking about?

This is OCUK Graphics sub it's almost entirely conjecture and quite often it's put up by people who want to make it seem negative.

essentially this is a fanboy hype thread. AMD fanboys predicting failure and nvidia fanboys praising nvidia for something.

 

[bru]

Uh oh, is this 480gtx all over again, randomly announce/launch it early yet only actually promise real supply for Q1 the next year and for a more expensive market and even then it's a salvaged and not a full part.

Looking it up the Quadro 5000 had another SM disabled compared to the 480gtx but had higher clocks, 225W TDP compared to the 480gtx 250W, keep in mind that at the time the 5870 was a 188W card.

Nvidia announce the card, salvaged and already 300W and list actual availability 9 months later, really doesn't sound great. Figured they learned about large cores early on a process from 40nm as they avoided the problem at 28nm(1-2 process nodes after the rest of the industry knew not to do it).

Would Nvidia release an over 300W consumer card? 2 more SMs enabled but lower clocks, same or more SMs disabled, lower clocks and trying to hit 250W?

This is coming across as more and more of a stunt launch to beat AMD to the punch(considering they already announced Pascal as the first 3d memory card back in 2014 CES, it wouldn't be new) and they have to respin GP100 and we'll have GP110 by some point in 2017.

I do get the feeling that you might be right and it could be the 480 saga all over again.

I just hope that it isn't going to be like that and that it isn't a sign of what the initial 16/14nm stuff is going to be like. Smaller size chips could well be the much better way to go, it makes sense seeing that such a big chip being their first 16nm part surely cannot have particularly good yields.
This could stand AMD in good stead with the Polaris 10 and 11, of course we still don't really know much about them at all.
It is certainly going to be interesting over the next few months


On a completely separate note, looking at the core count.

GK110 2880
GM200 3072
GP100 3584 (3840 if Anandtech is right)

That is quite a sizeable jump in numbers.

Then look at the clock speeds

GK110 745MHz
GM200 948MHz
GP100 1328MHz

Seems again like a sizable increase

But performance doesn't seem to show that sort of progression, with the few numbers that they have given us. with the TDP at 300W as well, I'm not really sure what to make of it all at the moment, of course we have to assume the numbers are correct seeing as they are direct from NVidia themselves, something just seems a little off.