Rumors are going around on the web that NVIDIA's GM204 "Maxwell" based GTX 800 graphics cards are slated for a late September release. Word on the street is that the first generation will be based on the 28nm process, just like the current generation of graphics cards.
According to SemiAccurate, NVIDIA is preparing at least four GeForce SKUs to be released before jumping into smaller fabrication process. The list includes:
GeForce GTX 880 Ti
GeForce GTX 880
GeForce GTX 870
GeForce GTX 860
All these cards will differ in Streaming Multiprocessor count. Various sources claim that full GM204 chip would feature between 15 and 20 SMs (1920-2560 CUDAs). Maxwell core is obviously much more power efficient than Kepler, thus the next flagship will most likely have less cores than full GK110 processor. This has not happened before.
The surprising news is that NVIDIA reportedly plans a die-shrink approximately four to six months after the launch of Maxwell and that the company plans on jumping straight to the 16nm process! This means the company will skip the 20nm process and will be first to adopt TSMC's 16nm FinFET technology.
There is some confusion though as to how NVIDIA will market the 16nm parts, the source claims the GM204B cards will likely carry the GTX 8xx naming scheme, perhaps with a special postfix to the naming scheme to differentiate these cards from the 28nm versions. Alternatively, if TSMC's 16nm yields aren't as good as expected and the GM204B gets pushed back it's possible that the die-shrink may be launched as the GeForce 900 series.