AIDA64 Accidentally Confirms GP102 and GP107 GPUs

[awesomeness10120]

Quote

The second batch of chips is for the famed GP102 core. This is the very same core that will be powering the next GTX 1080 Ti and the next TITAN series of graphics cards from Nvidia. We expect the die to be around 472mm in size and will probably have around 3840 cores. You can read up more on the GP102 and GTX 1080 Ti editorial over here. The device IDs are divided into two variants but are part of the same family, so it is possible that this particular Pascal GPU only has one chip configuration.

So now we have more solid evidence for another GDDR5/X powered high end Nvidia die. Guess GP100 (and maybe Vega) don't have very good yields.

Quote

Near the bottom of the barrel we have the GP107 and GP108 dies. The GP107 will be powering the GTX 1050 series of GPUs and several mobility variants. This is the low budget gaming spectrum of Nvidia GPUs and will probably be focused on OEM sales. The GP108 is the GPU that will only be useful (practically speaking) for HTPC purposes and not any amount of serious gaming.

Sources: 

http://wccftech.com/nvidia-pascal-device-id-gp-102-gpu-confirmed/

Very interested to see when HBM2 will make it to consumers.

[Enderman]

what made you think the 1080ti will have GDDR5X?

all ive heard is that it will use HBM2

[SansVarnic]

Interesting

[awesomeness10120]

19 minutes ago, Enderman said:

what made you think the 1080ti will have GDDR5X?

all ive heard is that it will use HBM2

Why would Nvidia make two HBM2 GPUs? Wouldn't that make manufacturing of P100s and Geforce cards more and more expensive? GP102 with GDDR5X would alleviate HBM2 supply problems but at the same time allow for an "enthusiast" Nvidia product.

[AlexGoesHigh]

31 minutes ago, Enderman said:

what made you think the 1080ti will have GDDR5X?

all ive heard is that it will use HBM2

This is not the first time the GP102 sees light, someone last week posted a chart containing the specs of that chips saying it was going GDDR5/X memory for the new titan and 1080ti which it kinda make sense since nvidia said the tesla P100 would have a staged release and the last faze would be general availability by early 2017 and before that it would only sell P100 in massive bulk orders to big biz and on their server enclosure thingy that i don't remember the name of.

 

That last week post about new titan/1080ti said it would come later this year, my guess is that HBM2 for consumers (at least for nvidia chips) is coming mid 2017 with the 1100 series since by them yield should be high enough for mass production for consumer markets.

[awesomeness10120]

9 minutes ago, Sampsy said:

-snip-

I don't know about lack of competition, as AMD is launching new architectures this year and in 2017, but I can see Nvidia being overconfident. GP100 is definitely having bad yields, (their top end multi thousand dollar Tesla P100 can't even have all cores enabled on its GPU) but I only hope that GP102 and GP100 make it to consumers so GP104 can get a price drop

 

[VagabondWraith]

I see this going the same way as GTX 780 > OG Titan > 780 Ti > Titan Black. GTX 1080 Ti (Cut-down GP102 GDDR5X) > Pascal Titan (Full GP102 GDDR5X) > GTX 1180 Ti (Cut-down GP100 HBM2) > Pascal Titan Ultra (Full GP100 HBM2). Then nVidia's Tock process Volta.

[Curufinwe_wins]

2 minutes ago, VagabondWraith said:

I see this going the same way as GTX 780 > OG Titan > 780 Ti > Titan Black. GTX 1080 Ti (Cut-down GP102 GDDR5X) > Pascal Titan (Full GP102 GDDR5X) > GTX 1180 Ti (Cut-down GP100 HBM2) > Pascal Titan Ultra (Full GP100 HBM2). Then nVidia's Tock process Volta.

None of that is correct....

 

The 780ti and Titan Black are identical in gaming (provided thermal limits are removed and clock speeds are the same), as are the 780 and Titan. They all share the same core.

 

Additionally, the P100 and GP102 dies are identical according to all the links (3840 cuda cores etc etc), so it is likely GP102 is the 780ti variant of the P100 "Titan". 

[awesomeness10120]

3 minutes ago, Curufinwe_wins said:

None of that is correct....

 

The 780ti and Titan Black are identical in gaming (provided thermal limits are removed and clock speeds are the same), as are the 780 and Titan. They all share the same core.

 

Additionally, the P100 and GP102 dies are identical according to all the links (3840 cuda cores etc etc), so it is likely GP102 is the 780ti variant of the P100 "Titan". 

GP102 is now real, but there are some inherent differences between GP100 and GP102. For one, FP64 performance is completely gimped on GP102 to improve costs, as gaming doesn't use FP64. Also, GP102 doesn't use HBM2.

[Curufinwe_wins]

Just now, awesomeness10120 said:

GP102 is now real, but there are some inherent differences between GP100 and GP102. For one, FP64 performance is completely gimped on GP102 to improve costs, as gaming doesn't use FP64. Also, GP102 doesn't use HBM2.

1. we don't know if it uses or doesn't use HBM2 (I doubt very much it is using GDDR5x just saying) This leak doesn't confirm ANYTHING ABOUT THE CORE CONFIG OR MEMORY SUBSET.... We are still going off the previous assumption/leaks. Read the actual leak...

 

2. That is why I called it the 780 ti compared to Titan black. Exact same in everything other than DP perf.

[Curufinwe_wins]

Seriously... this leak CONTAINS NOTHING OTHER THAN DEVICE ID's! IT DOESN'T SAY A SINGLE GODDAMN THING ABOUT CORE CONFIGs, MEMORY STRUCTURE, or EVEN DIE SIZE.

 

All of the other data in the WCCF-tech article is still the 100% unsubstantiated rumor mill.

[Trixanity]

They're definitely having yield issues. I mean they probably already have at least minor issues with gp104 and having to invent a "Founder's Edition" while waiting on more chips to be ready to satisfy the market. 

 

And it's easy to see that a 600mm die will be hell to produce in a large amount. So (G)P100 will be their only behemoth for now. I'm quite sure they prefer to sell whatever chips they get for enterprise customers, where the money is better. 

 

It's only natural that a smaller GP102 is made to fill in while everything gets up to speed. The yields will be better, albeit probably not too good either.

[Carclis]

2 hours ago, Sampsy said:

Not a happy bunny but not very surprising either - this trend has been going on for a while. If the 1080ti is on gp102 rather than gp100 it really has nothing to do with yields and everything to do with a lack of competition. Nvidia has been matching AMD's high end offerings with what is really their mid end for a while now. The 980 was never the high end maxwell chip, it was just marketed as such when it was released, eventually to be replaced with the real high end in the form of the 980ti and titan x.

 

This move would just be another iteration on the same thing. Market the mid end 1080 as a high end chip and charge whatever people are willing to pay to have the fastest card around. Then do the same for a mid-high end gp102 1080ti, bumping down the 1080. Finally release the full fat gp100 and yet again Nvidia manage to get people to pay the very high price to have the fastest card available. Because AMD aren't managing to match Nvidia in a timely fashion Nvidia can get away with charging premium prices for what are really non-premium products. Why would they sell a mid end chip for a mid end price if AMD have no response to it - they may as well charge what the market will bear.

 

Sure you could argue that ultimately the 1080 is still a very fast card and it's purely price/performance that should matter to consumers, not how much it costs Nvidia to make it or where it sits in their product portfolio. But it's frustrating as hell knowing the huge profit margin they are making at our expense, drip feeding consumers just enough to keep buying their cards while spending as little as they can get away with on R&D. It's completely due to a lack of competition from AMD and it's only going to lead to GPU technology falling far behind where it could be. I wouldn't be surprised if Nvidia do something similar to intel and start slowing down their architecture development - why bother bringing out Volta when you can keep selling Pascal for the next x years?

 

That is more or less just Nvidia being Nvidia. Big chips are expensive to produce and there is no need to produce them when most of AMD's focus is on the mid-low end this generation. Don't confuse this with AMD not being able to match Nvidia. Remember, the mid and low end are where the majority of money is made because yields are so much higher along with sales too, plus this is the same area that AMD is focusing on for consoles too which are all on AMD hardware. As you can probably see both companies focus on opposite ends of the market, but they have other offerings which keep prices in check. If you want exorbitant prices and stagnation, look at Intel.

[zMeul]

4 hours ago, awesomeness10120 said:

Why would Nvidia make two HBM2 GPUs? Wouldn't that make manufacturing of P100s and Geforce cards more and more expensive? GP102 with GDDR5X would alleviate HBM2 supply problems but at the same time allow for an "enthusiast" Nvidia product.

because, from a previous rumor, the GeForce Pascals do not have FP64 units - that means nVidia has two manufacturing lines, one for Tesla / and possibly Quadro  and one for GeForce / Quadro