GPU Pascale GPU spotted, could launch in June

[Ceros_X]

http://wccftech.com/nvidia-pascal-graphics-cards-spotted/

Didn't see this linked here, key thing I spotted in this article was the following image (not really addressed in the article):

1/3rd size of PCI-E Card and a Maximum TDP of 300W coupled with 2x performance per watt estimated improvement over Maxwell could lead to some awesome SFF build possibilities!

(Admin if you want to move/reformat this to news please do, didn't want to post up something crappy looking there)

 

[jeshikat]

This form factor has been shown before. It was at another conference sometime last year.

The idea is cool but I don't think it'll be of much interest to the SFF community because these cards have mezzanine connectors on the bottom and so require special motherboards to use.

 

[QinX]

By definition this product is NOT a graphics card or GPU for that matter.
This product is solely aimed at HPC or High Performance Computing and is used for mathematical calculations.

the Pascal GPU that will be based on this is still a long way out. First we will be getting GP104 which is the smaller core that is based on GDDR5X.

 

[PNP]

>using wccftech as a source
Oh dear, oh dear.

That's an NVLink card designed for supercomputers to use as a co-processor. Yes, the size is is something to marvel at, but given that the TDP has been revealed to be 300 W, a decent cooling solution in the consumer space would balloon the size. Also, for a card that's shipping in June/July, there's been a surprising lack of real die shots. Among other things, that inter-poser is too dull to be the real deal:

Spoiler: Slide from GTC courtesy of Anandtech

(We need spoiler tags to hide images, btw)

I don't expect a consumer GP100 card with HBM2 until Q3 2017 and even then I expect a supply crisis like what we saw with the Fury X.

 

[jeshikat]

(We need spoiler tags to hide images, btw)

There are, just click the little + sign to the right of the video tags for them.

Spoiler: Yo Dawg

 

[PNP]

Spoiler: Yo Dawg

Y-you monster.

 

[Ceros_X]

>using wccftech as a source
Oh dear, oh dear.

Are they the Inquirer of Tech Reporting or something?

 

[PNP]

Are they the Inquirer of Tech Reporting or something?

Perhaps you meant Enquirer? The Inquirer is a tech reporting site too. I couldn't tell you since I don't regularly read either of them, but I can tell you that Wccftech is a rumor mill aggregate that does not particularly care for the reputation of its sources. It's a fun read on a slow news day, but I'd never cite it.

 

[Ceros_X]

Perhaps you meant Enquirer? The Inquirer is a tech reporting site too. I couldn't tell you since I don't regularly read either of them, but I can tell you that Wccftech is a rumor mill aggregate that does not particularly care for the reputation of its sources. It's a fun read on a slow news day, but I'd never cite it.

I meant Enquirer - I don't read it as you may guess. Thanks for the heads up, I will avoid them in the future.

 

[jØrd]

http://www.gizmodo.com.au/2016/04/meet-the-dgx-1-nvidias-stupidly-powerful-supercomputer/

 

[EdZ]

I think we might see very limited antiquates of a GP100-based GPU sometime this year. Tesla P100 is using a MASSIVE die on a brand new process, and already with very tight binning (56 of 60 SMs enabled). Nvidia will be building up a large pile of not-quite-good-enough GP100 dies during P100 production, and I can't see them passing up the change to sell them. Maybe as a Titan rather than a GTX, but even if the chip was binned down to 44 SMs (16 disabled, or a over quarter of the chip!) it would still have as many SMs as a 980ti, but at a higher clock speed, with the added memory (and memory bandwidth) of HBM2, AND the arch improvements of Pascal.

 

[QinX]

The Tesla P100 already has a much higher clockspeed considering it's a enterprise grade chip. Running at a baseclock of 1328MHz and a boost of 1480MHz.
By comparison the GM200 based Tesla M40 only has a baseclock of 948MHz and a boost of 1114MHz.

If the same ratio applies to a GTX 1080 Ti then we could be looking at a base clock of 1400Mhz. That alone already gives a potential 40% boost in performance.

 

[BirdofPrey]

I think we might see very limited antiquates of a GP100-based GPU sometime this year. Tesla P100 is using a MASSIVE die on a brand new process, and already with very tight binning (56 of 60 SMs enabled). Nvidia will be building up a large pile of not-quite-good-enough GP100 dies during P100 production, and I can't see them passing up the change to sell them. Maybe as a Titan rather than a GTX, but even if the chip was binned down to 44 SMs (16 disabled, or a over quarter of the chip!) it would still have as many SMs as a 980ti, but at a higher clock speed, with the added memory (and memory bandwidth) of HBM2, AND the arch improvements of Pascal.

Would they be more likely to bin them as Titans or Quadros though?

 

[michaelmitchell]

Found this recently for those interested
http://wccftech.com/nvidia-pascal-gp104-gpu-gddr5x-leak/

 

[FCase]

Found this recently for those interested
http://wccftech.com/nvidia-pascal-gp104-gpu-gddr5x-leak/

Pascal die may have shrunk but it does not look like the card diminished in size at all.

Mal

 

[GuilleAcoustic]

Pascal die may have shrunk but it does not look like the card diminished in size at all.

Mal

I really hope the dual 8 pins PEG is a joke

 

[QinX]

I really hope the dual 8 pins PEG is a joke

GTX1080 is likely to be a 2x 6-pin or single 8-pin chip. with the aftermarket uberclocking custom PCBs having dual 8-pin for LN2 sessions.

 

[Phuncz]

That's an "MSI Gaming" GTX 980 card with the core area of the so-called MSI Gaming GTX 1080 photoshopped on it, so it's a custom PCB. Those dual 8-pin connectors and 30% of the board size is not reference design.

 

[EdZ]

The proposed '1070/1080' still use GDDR memory, so they still need enough PCB area for the lanes to fanout. Could end up as a more compact PCB like the GTX 670 (~175mm long), but the limiting factor will be TDP and thus heatsink size. Put an undersized heatsink on (like the R9 Nano) and you run into thermal limits and have to throttle back. The current rumour is that like the GTX 970, the GTX '1070' will not be issued with a reference design, so there's the possibility some OEMs may produce a short-PCB-big-cooler version.

 

[Phuncz]

The R9 Nano's throttling is being blamed on it's much lower TDP restrictions, apparently. But I see what you mean, if it was a 250W card, it would rarely be able to deliver stable performance.

Although we can be hopeful for the 14/16nm generation having very low TDPs.