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

[Huggie86]

Here is first ever picture of either GTX 1080 Ti, 1080X or could be 1090 GP104-400 full fat chip with GDDR5X leaked that will replace 980 Ti.



http://wccftech.com/nvidia-pascal-gp104-gpu-gddr5x-leak/

These GDDR5X chips on GP104-400 PCB is odd, the part number 6GA77 Z9TXT was not listed on Micron site.

So are these definitely looking like GDDR5X chips then?

 

[Mtom]

Citation is common sense. Unless nvidia decided to throw money at 2 different pascal designs, I don't see info of Tesla pascal being unrelated to consumer parts.



AMD looks like is releasing mid range 390 replacement without touching fury x. At least it seems so. I don't see why it would be commercial suicide for nvidia to do similar thing to replace high volume parts with new ones containing smaller fooprint in order to start generating some revenues.

Also, I am very curious of AMD Press conference the other day when they sounded so sure Polaris will be extremely competitive, and that they are in much better competitive position that they haven't been in years. Somehow that doesn't compute, unless they are just sweet talking the share holders

Why it doesn't compute for you? Each card has to be competitive in its on segment against its counterpart.

 

[bru]

Ever wonder how they problem solve chips.

Good video, it is really good to get this sort of access that we wouldn't normally get, well worth a watch.

 

[muziqaz]

Why it doesn't compute for you? Each card has to be competitive in its on segment against its counterpart.

It doesn't compute on a lot of levels. For one AMD always spins their so called advantage, which rare is advantage anyway. They were so confident that they have head start on release date with their finfet products by quarters, yet we have more chatter about nvidia imminent release than AMD. Obviously that chatter mostly is echo chamber, but still. Also chatter is saying 1070/1080 is about to be released/revealed, which should be at least a bit faster than current top end, while AMD was demoing their polaris 10 with hitman performance at max on par with fury cards.

In my eyes, if AMD is so cocky about their new products, that would mean they would have been released by now and kicking nvidias products left and right.
What upsets me more is AMD attitude to situation as if saying that comparatively performing cards from both sides are good for AMD. Yeah they are, we all saw AMD latest financial reports. When all is equal in GPU market nvidia will still win with their PR and customer base

 

[Rroff]

Good video, it is really good to get this sort of access that we wouldn't normally get, well worth a watch.

nVidia has been opening up a lot of that stuff to the media/press in the last few weeks - not sure why i.e. the emulation lab which they've previously almost kept under wraps.

Pretty cool how they use Gallium cutting to get a side on view of a faulty transistor.

 

[muziqaz]

nVidia has been opening up a lot of that stuff to the media/press in the last few weeks - not sure why i.e. the emulation lab which they've previously almost kept under wraps.

Pretty cool how they use Gallium cutting to get a side on view of a faulty transistor.

The calendar on the wall dates 2012. Wonder if someone is very forgetful or video was shot in that year

 

[Rroff]

They were handing out GM200 failed cores and some of the hardware they are using didn't exist in 2012 so I suspect it isn't that old.

 

[ADT]

So is pascal worth the wait then over the current cards??

 

[pmc25]

So are these definitely looking like GDDR5X chips then?

Yes. But any gp104 sku with gddr5x can't come before Oct / Nov time. Mass production of gddr5x only starts in July.

I'd guess the GP104 is 280-295mm2. Polaris 10 is 230mm2. I suspect transistor counts will be higher in the AMD product though .... if GP100 is anything to go by, NVIDIA have density problems.

 

[bru]

The calendar on the wall dates 2012. Wonder if someone is very forgetful or video was shot in that year

It was filmed at GTC this year when NVidia invited Blunty out to silicon valley.

 

[Mtom]

It doesn't compute on a lot of levels. For one AMD always spins their so called advantage, which rare is advantage anyway. They were so confident that they have head start on release date with their finfet products by quarters, yet we have more chatter about nvidia imminent release than AMD. Obviously that chatter mostly is echo chamber, but still. Also chatter is saying 1070/1080 is about to be released/revealed, which should be at least a bit faster than current top end, while AMD was demoing their polaris 10 with hitman performance at max on par with fury cards.

In my eyes, if AMD is so cocky about their new products, that would mean they would have been released by now and kicking nvidias products left and right.
What upsets me more is AMD attitude to situation as if saying that comparatively performing cards from both sides are good for AMD. Yeah they are, we all saw AMD latest financial reports. When all is equal in GPU market nvidia will still win with their PR and customer base

As i said each card is in his own segment. Polaris 10 as the rumors said is just the 480 card, the mid range one, with your logic it needs to be faster than the 380. But according to the leaks its faster than the 390(x)...so how it is not competitive? The one needs to be faster than Fury is next step up card which will be the 490 is the number scheme remains.

You are right about NV PR and customer base. There was a good video lately showing that when they don't have anything to show for 1/2-3/4 years against AMD in the DX10 era, they still didn't lost market share. They are like apple...(more than)half their customers buy NV card because its the cool thing.

 

[D.P.]

Yes. But any gp104 sku with gddr5x can't come before Oct / Nov time. Mass production of gddr5x only starts in July.

I'd guess the GP104 is 280-295mm2. Polaris 10 is 230mm2. I suspect transistor counts will be higher in the AMD product though .... if GP100 is anything to go by, NVIDIA have density problems.

Mass availability of GDDR5X is scheduled for this summer, well ahead of schedule. Still not early enough for a June launch GPU of course but I wouldn't be surprised to see a GDDR5 GPU in august or sepember


I wouldnt use the GP100 transistor density as any indicato on for lower end parts.It is a well known practice to include dark silicon that improves thermal ad electrical charcteriscs, reduces defect rates etc. The GP100 has a huge profit margin, thus they can afford to make a bigger GPU if it has advantages to allow higher clock speeds, or longer life expectancy, etc.

 

[hominid]

I'd guess the GP104 is 280-295mm2. Polaris 10 is 230mm2. I suspect transistor counts will be higher in the AMD product though .... if GP100 is anything to go by, NVIDIA have density problems.

Don't forget they are on entirely different processes, NV on 16nm and AMD on 14nm - and both will have different rules for transistor spacing etc

AMD have a 2nm advantage over NV, so 2/16(NV process)*100 = 12.5% smaller process for AMD.. so their 232nm is equivalent to NV 232*12.5 = 261mm squared in, very, simplistic terms.

 

[Harlequin]

Micron are not in production of gddr5x yet the process for there product is sample > contact factory > production. Currently only the slowest gddr5x is at contact factory the rest is still at sample

 

[bru]

Yes. But any gp104 sku with gddr5x can't come before Oct / Nov time. Mass production of gddr5x only starts in July.

I'd guess the GP104 is 280-295mm2. Polaris 10 is 230mm2. I suspect transistor counts will be higher in the AMD product though .... if GP100 is anything to go by, NVIDIA have density problems.

According to the latest leaked shots the GP104 is 333mm^, I very much doubt that the transistor counts will be higher in the AMD Polaris 10, probably nowhere near it seeing as Polaris 10 is just over two thirds the size.

 

[andybird123]

Don't forget they are on entirely different processes, NV on 16nm and AMD on 14nm - and both will have different rules for transistor spacing etc

AMD have a 2nm advantage over NV, so 2/16(NV process)*100 = 12.5% smaller process for AMD.. so their 232nm is equivalent to NV 232*12.5 = 261mm squared in, very, simplistic terms.

It doesnt quite work like that. Both samsung's 14nm and tsmc's 16nm ff+ are based on 20nm with finfets. There are minor differences between the two but it's not a straight 14 vs 16 as a percentage - there isn't even that much of a difference in pure size terms.

 

[Rroff]

AMD have a 2nm advantage over NV, so 2/16(NV process)*100 = 12.5% smaller process for AMD.. so their 232nm is equivalent to NV 232*12.5 = 261mm squared in, very, simplistic terms.

As Andy said ultimately they are both based off 20nm planar - dunno if it still holds true but with earlier products TSMC 16nm versus Samsung 14nm for the same design resulted in very slightly bigger cores on TSMC but they were also slightly more power efficient - last I heard they were getting either ~3% more power efficiency or ~6% higher clock speed (dunno if that is due to anything like compensating for issues like leakage or stuff like someone else mentioned with leaving areas of "dark silicon" to aid in thermals, etc. in layout).

 

[bru]

As Andy said ultimately they are both based off 20nm planar - dunno if it still holds true but with earlier products TSMC 16nm versus Samsung 14nm for the same design resulted in very slightly bigger cores on TSMC but they were also slightly more power efficient - last I heard they were getting either ~3% more power efficiency or ~6% higher clock speed (dunno if that is due to anything like compensating for issues like leakage or stuff like someone else mentioned with leaving areas of "dark silicon" to aid in thermals, etc. in layout).

Interesting.

If the die sizes mentioned earlier are correct, Polaris 10 (232mm^) replacing the Tonga chip (359mm^) is 64% of the formers size. Whereas on the NVidia front, the GP104 (333mm^ replacing the GM104 (398mm^) is 83.6% smaller.
So NVidia have effectively gone for a bigger chip rather than AMD's decision to go for a smaller one, by the sounds of it, the Polaris 10 real rival would be the GP106, which we have no idea what size it will be, but using a similar methodology could be 190mm^ ish (84% of the current GM106 at 227mm^) so AMD would actually have the bigger chips.

MMmmmmm

Confused, yup so am I.

 

[D.P.]

Don't forget they are on entirely different processes, NV on 16nm and AMD on 14nm - and both will have different rules for transistor spacing etc

AMD have a 2nm advantage over NV, so 2/16(NV process)*100 = 12.5% smaller process for AMD.. so their 232nm is equivalent to NV 232*12.5 = 261mm squared in, very, simplistic terms.

Both the 14 and 16nm proceesses are actually the same 20mm nodes. These are just marketing names.

 

[Squink]

There are minor differences between the two but it's not a straight 14 vs 16 as a percentage - there isn't even that much of a difference in pure size terms.

Yeah - a fraction of a millimetre at best