Yep, AMD could have "launched" last December as well. The market share will tell the true story of who has cards and who doesn't.
The more I think about it, the more I would be inclined to paper launch the desktop parts and get mobile out ASAP. Don't expect AMD to start a price war on mobile but they do have a much smaller chip (232 mm2 vs 300 mm2) and architecturally there isn't much difference between Polaris 10 and GP104. The top end clock speed advantage Pascal seems to have won't really be a factor in mobile.
That really isn't paper launch though. The post specifically says:
In other words, they don't know if Nvidia is simply having a press event to let the tech review sites early access to the cards so they can start writing their articles to be available to post them on the date of the actual launch.
We see this kind of thing all the time on next generation items which have large changes over previous generations. A lot of the information is made available in bits and pieces before the product launch to let people have a chance to digest the implications of the new product and drum up early product demand. On complex items it can take days or weeks to go over all the new features and the benefits of those features. By getting some of that information out early, it gives the consumers and the reviews a chance to think about the ramifications and generate better questions and test conditions to see such benefits and drawbacks of the design decisions that were made for the product.
I mean, by your definition you are saying that no one can release any information about a product until the day the product is on the shelves, which goes against almost all marketing doctrine ever created. It would be like saying there is no need to advertise that a movie is coming out until the day the movie is in the theaters....
That really isn't paper launch though. The post specifically says:
In other words, they don't know if Nvidia is simply having a press event to let the tech review sites early access to the cards so they can start writing their articles to be available to post them on the date of the actual launch.
I mean, by your definition you are saying that no one can release any information about a product until the day the product is on the shelves, which goes against almost all marketing doctrine ever created. It would be like saying there is no need to advertise that a movie is coming out until the day the movie is in the theaters....
If you believe the latest rumour by DigiTimes, the opposite could happen:
www.digitimes.com/news/a20160408PD205.html
Paper-launching seems very unprofessional. IIRC, Maxwell2 was announced somewhere in August 2014, and three weeks later the products were in store. En masse. I'm sure Nvidia can repeat this format with Pascal again.
So why people were talking about 7970 being a paper launch when it was announced on December 22, 2011 and it was in stores on January 9, 2012. That's three weeks.
And I don't know about the rest of the world but I got 2 of them at launch day.
Soo, maybe this what we call double standards :whiste:
What do you think are the chances GP104 could actually have 3200 (50x64) shaders instead of rumored 2560 (40x64)? At least the rumored top SKU - 1080Ti?
GP106 is rumored to be right at 200mm2, which puts it in between Polaris 10 and 11, but closer to Polaris 10. Given the supposed die size of both GP104 and GP106, GP106 is probably around 60% the cores of GP104, but with a question mark on bus size (GK106 was 192-bit while GM206 was 128-bit).
We know that the diesize should be around 300mm², no way you can fit 50 Clusters in there. 40 sounds the most logical and then 60 for GP102.
Been some time since i checked this thread,any good chance of a $200-$250 Pascal this year?
What do you think are the chances GP104 could actually have 3200 (50x64) shaders instead of rumored 2560 (40x64)? At least the rumored top SKU - 1080Ti?
I assume people think its 40, because its natural to cut 1/3 from that Tesla p100 diagram...but its only a diagram after all. And we had non-symmetrical diagrams/layouts to older GPUs before, like for example OG Titan:
GP106 in general is a mystery at this point, but I think you can be assured there won't be any desktop parts any time soon. There's a rumor that nVidia might do a really cut down GP104 that has 970-like performance but it's for OEMs only.
The chip itself can't be too late, with it (I think?) being in that automatic car thing that was shipping in Autumn or something.
Not entirely impossible.
If we assume that a Pascal SM without the FP64 units is half the size of a Maxwell SM (since it only has 64 CUDA cores vs. 128 in Maxwell), then you would be able to fit in twice as many SMs per mm2 (which would of course result in the same number of CUDA cores).
So a 600mm2 Pascal GPU on 28nm would have 48 SMs (twice that of GM200 with 24). Make the transition to 16nm and you would roughly half the die size down to 300mm2, which is more or less what GP104 is rumored at. 48 Pascal SMs is equal to 3072 CUDA cores, not quite 3200, but close.
Using the same logic a 2560 CUDA core GP104 (40 SMs) would be about 250mm2, and a 3200 CUDA core GP104 would be about 313mm2. So it basically comes down to the eventual die size of GP104. Both 250mm2 and 313mm2 are within the realm of possibility imho.
Well, i hope for a 3200 CC version then. It would suck to buy a card with lesser number of cores than 980Ti/Titan X, even if it was hypothetically / most likely somewhat faster due to higher clocks and IPC improvements. The fact its already most likely going to have less VRAM than TX is enough of a downer. I guess its unlikely we could see more than 8GB VRAM on any of the assumed 104 SKUs, right? As far as my needs are concerned 12/16GB GDDR5 > 8GB DDR5X.
But that's exactly what happened with 780 Ti to 980 Ti. If these guys didn't even publish CU numbers we wouldn't even know the difference, so I think its pointless to care about that metric.
I dont think we really "know" that, not to mention the GP104 cluster may not be the exactly the same as GP100 cluster. I mean, it can have the same 64 FP32 cores, but not those additional 32 FP64 units (but say only 8 instead per cluster)...that should save loads of die space, right?
Admittedly, what u say make sense (40 for 104, 60 for 102), but thats under assumption there actually is GP102 in the making. What if it is not and its just a rumor?
But that's exactly what happened with 780 Ti to 980 Ti. If these guys didn't even publish CU numbers we wouldn't even know the difference, so I think its pointless to care about that metric. I sort of understand the VRAM issue but again, if it performs due to better memory management, does it reall ymatter?
Of course it will save space to cut the HPC features, but architecture tweaks like the better preemption cost you space, the new SM layout also will need more diesize, but should help per shader performance.. Transistor Density goes up 2x, but if this chip is really ~300 mm²(yes we don't know it, but it seems so after the leaked pic) then there is no way to get 3200 shader of GP100 has just so few
GP102 is in nvidias drivers, that's more credible than most of the rumors which are floating around. I'm more sceptical of this rumored GTX1080Ti and believe still in 2 GPUs at the moment, one of them with GDDR5X shipping in June, as Micron stated partners are shipping GDDR5X products in their Q2.