Image source: NVIDIA.
Whenever graphics specialist NVIDIA (NVDA 3.17%) develops a graphics architecture, it builds multiple chips based on that same fundamental architecture. The broad array of chips is intended to serve a diverse range of performance/power/cost requirements, as a single chip isn't going to cut it.
Thanks to the folks at FinalWire (via VideoCardz.com), creators of the AIDA64 system diagnostic tool, we now know exactly how many Pascal chips NVIDIA has planned.
The answer is six
According to the post, there will be no fewer than six distinct Pascal-architecture based graphics processors. These are codenamed GP100, GP102, GP104, GP106, GP107, and GP108.
GP100 is already announced and it is the chip that powers NVIDIA's Tesla P100 compute accelerator. GP104, too, has been announced and powers the recently announced GeForce GTX 1080 and GeForce GTX 1070 graphics cards targeted at high-performance gaming desktops (and, likely, at high-performance gaming laptops as well).
This leaves GP102, GP106, GP107, and GP108 still to be revealed. However, based on their code names, we can make some educated guesses as to how they will ultimately be productized.
GP106 and GP107 replace GM206 and GM107
The first product based on NVIDIA's Maxwell architecture was a part known as GM107, marketed as the GeForce GTX 750 Ti. NVIDIA never released a GM207 (i.e., a second-generation Maxwell part for the entry level of the gaming market), instead relying on a partially disabled GM206 to serve as the GeForce GTX 950 -- the virtual successor to the GTX 750 Ti.
It would seem that during the Pascal generation, NVIDIA will introduce proper successors for both GM206 as well as GM107. I wouldn't be surprised to see the GP106 power the GTX 1060 exclusively (more on what GP106/GTX 1060 may be like here), with the lower-end GTX 1050 packing a cheaper-to-make GP107 part.
GP108 replaces GM108 for low-end applications
Although GM108 doesn't appear to have been used in any gaming-oriented desktop graphics cards, it was used as an entry-level discrete graphics chip for notebooks. I would expect a similar usage for GP108 as its predecessor: laptops that are, at best, used for light gaming. From a revenue perspective, I would imagine that sales of such chips don't get recognized under gaming but rather under the declining OEM and IP business.
GP102 as the new Titan
Right now, NVIDIA actually has it pretty good with the GP104; NVIDIA's own Founders Edition GeForce GTX 1080 card based on the full GP104 will retail for $699. However, since the GeForce GTX 1080 is much faster than NVIDIA's previous $999 graphics card, the Titan X, the company will need a faster chip in order to be able to sell cards at higher price points.
That's what GP102's job is likely to be.
There have been leaks suggesting that GP102 is essentially the GP104-plus-50%, which is sensible considering that's what GM200 was relative to GM204.
NVIDIA says all Pascal chips "taped out," so rollout likely won't take too long
According to NVIDIA CEO Jen-Hsun Huang on the company's most recent earnings call, all of the chips based on the Pascal architecture have been "taped out." This means that the designs are done (though there may be fixes required that force NVIDIA to do one or more iterations of the silicon).
Considering that the most complex chip (GP100) as well as the third-most-complex chip (GP104) are in production today, I would be surprised if the company were having issues with the smaller and easier-to-build GP106/GP107/GP108 chips.
GP102 is probably going to be the trickiest of the bunch. Although NVIDIA will be able to charge a pretty penny for GP102, it will need to be sold at gamer-friendly prices. This is in stark contrast to the GP100, which is aimed at hyper-scale/high-performance computing with very deep pockets that will pay big as long as a product reduces their overall total cost of ownership.
So, the key for GP102 will be to make sure that chip yields are high. And, if GP102 uses the more exotic HBM2 memory rather than the GDDR5X that GP104 uses (though one rumor points to the use of GDDR5X), then NVIDIA is going to need to be sure that this new memory type can be sourced in high quantities and at reasonable prices.
