NVIDIA Pascal Thread

[Silverforce11]

New unproven node, massive GP100, big orders from supercomputer builders who are waiting on these Teslas...

Ofc GP100 Tesla are shipped first, no doubts about that at all. Yields can be horrible, $6K or more each will make it worthwhile. Contract obligations need to be fulfilled first and foremost.

Q3 availability for GP104 is about right, mid-range 256 bit bus will need GDDR5X and that's not volume production until June.

They can launch GDDR5 SKUs earlier than that though. So GP106/107 parts could go early, no reason to delay those parts.

 

[tweakboy]

Very cool for hardcore programmers... good job.. gl

 

[jpiniero]

Yeah I kind of think it is fake. I don't expect GP104 to be faster than the 980 Ti but it probably would be close so I don't know how they would spin it unless they have something higher up in the chain then.

This doesn't seem accurate. The Tesla GPUs are supposed to have 32GB HBM2, but HyniX/Samsung are only producing 4GB stacks (16GB max) right now which leads me to believe the first GP100 chip released will be the Titan, not Tesla.

The Tesla is rumored to be a dual die product though. They could easily sell a cut GP100 as the first Titan and then the full GP100 later when yields get better.

 

[Silverforce11]

@jpiniero
How much faster is the 980Ti over the 980?

Now take the 980, x1.75 it for the node shrink (die size should return towards ~300mm2 for mid-range, 980 was a pretty big die) and it should easily beat the 980Ti.

 

[jpiniero]

How much faster is the 980Ti over the 980?

35-40%

Now take the 980, x1.75 it for the node shrink (die size should return towards ~300mm2 for mid-range, 980 was a pretty big die) and it should easily beat the 980Ti.

That sounds more like what the GP102 is.

 

[JDG1980]

Yeah I kind of think it is fake. I don't expect GP104 to be faster than the 980 Ti but it probably would be close so I don't know how they would spin it unless they have something higher up in the chain then.

I agree that the article looks fake and may well be based on the fan-made "roadmap" that was posted a couple weeks back. But I think that GP104, when it does arrive (October would be my best guess), is going to blow right past GTX 980 Ti. It won't be close.

Here is my rationale. First off, TSMC 16FF+ has about double the density of 28nm, so just a straight die-shrink of GM200 with no architectural changes whatsoever would give a ~300mm^2 chip with the performance of the Titan X. But the GM200 has a 384-bit memory controller. With GDDR5X, the width can be cut to 256-bit while still providing an increase in memory bandwidth (~20-25%). Then the extra space freed up can be used to add more shaders and ROPs. It wouldn't be at all out of the question for GP104 to have 4096 shaders and 128 ROPs, while keeping the die size below 350mm^2. Then you have to factor in clock speed improvements. FinFET helps a lot; the Apple A9 had a 32% maximum clock speed improvement over the planar Apple A8 in the same iPhone form factors. Add that to the increased shader count and memory bandwidth, and you could easily be looking at 50% more performance than even a mildly overclocked GTX 980 Ti. This is even assuming (which I do) that Pascal won't contain any architectural improvements over Maxwell that gamers should care about, except the new memory types.

 

[LTC8K6]

Nooooooooooooooooooooo!...Oh wait I could care less.

How much less could you care?

 

[Cookie Monster]

I think nVIDIA will push with the big Pascal regardless of yields because of the huge margins involved (in the GPGPU/HPC space) especially when they already have large orders to fulfil e.g the NOAA project. Not sure about a Titan version for the normal consumers, but definitely theyll aim to release its Tesla equivalent. The GK210 needs replacement.

On the otherhand, the timing of GDDR5X might indicate that the products might be launched later this year unless its going to use HBM2 or just plain GDDR5.

 

[jpiniero]

I agree that the article looks fake and may well be based on the fan-made "roadmap" that was posted a couple weeks back. But I think that GP104, when it does arrive (October would be my best guess), is going to blow right past GTX 980 Ti. It won't be close.

Assuming they keep the same price range for GP104, it'd be much smaller than 350 mm2. I'd say somewhere between 225-250 mm2. That's why I believe the GP102 rumors - the product you described sounds like it.

 

[tviceman]

Assuming they keep the same price range for GP104, it'd be much smaller than 350 mm2. I'd say somewhere between 225-250 mm2. That's why I believe the GP102 rumors - the product you described sounds like it.

Common sense dictates that the chip replacing GM204 as the new high end card won't be fast enough at 225mm2 to be the new high end. A 225-250mm2 chip may very well end up faster than the gtx 980, but it won't be faster than the 980 TI.

 

[jpiniero]

Common sense dictates that the chip replacing GM204 as the new high end card won't be fast enough at 225mm2 to be the new high end. A 225-250mm2 chip may very well end up faster than the gtx 980, but it won't be faster than the 980 TI.

Correct, that's why I think the roadmap is fake

 

[BTRY B 529th FA BN]

Jumping in for some mumbo jumbo trash

https://www.youtube.com/watch?v=kS4uOexbK5A

 

[JDG1980]

Assuming they keep the same price range for GP104, it'd be much smaller than 350 mm2. I'd say somewhere between 225-250 mm2. That's why I believe the GP102 rumors - the product you described sounds like it.

That seems way too small for a '4'-series chip. The smallest they've made so far was GK104, and even that was 294mm^2. A lot of people seem to be taking the FinFET FUD about high prices, low yields, etc. as gospel. But I'm much more skeptical about the doomsday predictions. We've all seen those charts showing per-transistor costs going way up under 28nm; one particular poster just loves to put them up again and again. But most of these charts are years old. In fact, there were some substantial optimizations to TSMC's 20nm and FinFET processes in 2014 that, according to Qualcomm, "The resulting technology is more cost effective compared to 28nm HKMG processes and is cost-competitive versus 28nm polySiON." We all know that the foundry FinFET processes saw long delays. But why? I think a big part of the reason is precisely that the foundries have been struggling to get costs down and yields up to the levels their customers demanded. Now, I believe, that time has finally come.

And it's certainly possible that Nvidia will hike the price for GP104. If they can beat the GTX 980 Ti ($649) by 50%, then why not charge $699? Or even $799? People will pay it.

I suppose we'll find out in six to nine months which of us is right.

 

[ShintaiDK]

That seems way too small for a '4'-series chip. The smallest they've made so far was GK104, and even that was 294mm^2. A lot of people seem to be taking the FinFET FUD about high prices, low yields, etc. as gospel. But I'm much more skeptical about the doomsday predictions. We've all seen those charts showing per-transistor costs going way up under 28nm; one particular poster just loves to put them up again and again. But most of these charts are years old. In fact, there were some substantial optimizations to TSMC's 20nm and FinFET processes in 2014 that, according to Qualcomm, "The resulting technology is more cost effective compared to 28nm HKMG processes and is cost-competitive versus 28nm polySiON." We all know that the foundry FinFET processes saw long delays. But why? I think a big part of the reason is precisely that the foundries have been struggling to get costs down and yields up to the levels their customers demanded. Now, I believe, that time has finally come.

And it's certainly possible that Nvidia will hike the price for GP104. If they can beat the GTX 980 Ti ($649) by 50%, then why not charge $699? Or even $799? People will pay it.

I suppose we'll find out in six to nine months which of us is right.

That's the benefit of selling a device rather than a component. The GPU cost itself is relatively low in the total cost. And it will be masked in the total device cost, even if the GPU itself cost 25-50% more to make. Then we may just pay 50$ more for a 500-600$ card. I think everyone got surprised how gamers are willing to pay. Record GTX and record K sales says it all.

 

[Bryf50]

That's the benefit of selling a device rather than a component. The GPU cost itself is relatively low in the total cost. And it will be masked in the total device cost, even if the GPU itself cost 25-50% more to make. Then we may just pay 50$ more for a 500-600$ card. I think everyone got surprised how gamers are willing to pay. Record GTX and record K sales says it all.

No it doesn't really say it all. Record GTX sales includes everything from the 950 to the 980 ti. From the looks of things the record sales are mainly due to the 970 and 960.

 

[nvgpu]

http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process

AMD is not even relevant for 2016, 1 small notebook GPU and 1 Pitcairn replacement 232mm2 GPU. Happy waiting for 2017, ADF.




Your post is thread crapping.




esquared
Anandtech Forum Director

 

[Kenmitch]

http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process

AMD is not even relevant for 2016, 1 small notebook GPU and 1 Pitcairn replacement 232mm2 GPU. Happy waiting for 2017, ADF.

???

Posted in wrong thread?

 

[tviceman]

That's the benefit of selling a device rather than a component. The GPU cost itself is relatively low in the total cost. And it will be masked in the total device cost, even if the GPU itself cost 25-50% more to make. Then we may just pay 50$ more for a 500-600$ card. I think everyone got surprised how gamers are willing to pay. Record GTX and record K sales says it all.

I was already under the impression you are giving that the cost of the chip relative to the total cost of the graphics card is not extraordinarily high. Do you know the ballpark price for current TSMC 28nm wafers? Does anyone here know?

Assuming they keep the same price range for GP104, it'd be much smaller than 350 mm2. I'd say somewhere between 225-250 mm2. That's why I believe the GP102 rumors - the product you described sounds like it.

GP102 would be an entirely new class of GPU within a given family of GPU's or would simply be a code name renaming from previous code names.

If it's an entirely new class, then it would mean GP100 would likely be exclusive to Tesla and Quadro and may have some graphical functionality stripped out of it to save on die space, while GP102 would be what GM200 is (focused on graphic and single precision). This seems unlikely to me. If it's a code name shuffle, kind of similar to what AMD did with product retail naming when the 6870 became a mid-range die and was slower then the 5870, then it will have branched off from Nvidia's last 3 generational code names. This also seems unlikely (although perfectly plausible). Furthermore, Kepler had rumors of a GK102 - which never manifested into reality.

Therefore I'm going with GP102 being a mythical nonexistent chip. While all the scenarios above are possible, it would be uncharacteristic at this point to make an additional GPU unless the costs of creating another large GPU and entirely limiting GP100's role consumer card make financial sense.

 

[raghu78]

???

Posted in wrong thread?

ignore him. he is an AMD basher. had nothing useful to contribute to Nvidia thread. so got started with his usual routine.

 

[jpiniero]

GP102 would be an entirely new class of GPU within a given family of GPU's or would simply be a code name renaming from previous code names.

It wouldn't be a new class, it would replace GM200 as the highest performer with gimped DP. GP100 would be the class above it, and would be priced as such ($1200+).

 

[ShintaiDK]

I was already under the impression you are giving that the cost of the chip relative to the total cost of the graphics card is not extraordinarily high. Do you know the ballpark price for current TSMC 28nm wafers? Does anyone here know?

~2500$

Lower end cards have much more problems with increased GPU cost than higher end due to proportions.

 

[Arachnotronic]

~2500$

Lower end cards have much more problems with increased GPU cost than higher end due to proportions.

$2500 might be TSMC's cost, but I'd imagine after TSMC gets its margins it's more like $5k.

 

[tviceman]

~2500$

Lower end cards have much more problems with increased GPU cost than higher end due to proportions.

$2500 might be TSMC's cost, but I'd imagine after TSMC gets its margins it's more like $5k.

ShintaiDK, can we get clarification? Is $2500 the cost to foundry customers or the manufacturing cost that TSMC incurs?

 

[tviceman]

It wouldn't be a new class, it would replace GM200 as the highest performer with gimped DP. GP100 would be the class above it, and would be priced as such ($1200+).

It would be a new class because Nvidia has never created a GPU to slot between it's flagship chipGM200/GK110/GF110 and it's performance chip GM204/GK104/GF114. It would be in no man's land performance wise because salvaged GP100 chips can easily slot into that if need be. GP102 doesn't and will never exist unless GP100 is not EVER going to get a consumer GTX release.

 

[Spanners]

http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process

AMD is not even relevant for 2016, 1 small notebook GPU and 1 Pitcairn replacement 232mm2 GPU. Happy waiting for 2017, ADF.

Ignoring that it's irrelevant to this thread you're still terrible at this, do you even read what you link?

"Project F is yet another codename for the chip designed at AMD HQ. What we’ve learned from previous LinkedIn leaks is that just because something is designed and reaches engineering state, it does not mean it will end up as a product for end-users. Project F could be one of such designs."

Even without that disclaimer there is no logical link between that article and what you wrote, nothing would imply that's the only chip/card they will or will not release this year. Great job as usual.