We all had thought that NVIDIA would have just one “big chip” based on the “Pascal” architecture, but it looks like there will be two. The first one is of course the GP100, which we were introduced to with the Tesla P100 HPC processor. The GP100 featured an extensive feature set with a combination of FP32 (single-precision) and FP64 (double-precision) CUDA cores. There are 3,840 SPFP and 1,920 DPFP giving you a total of 5760 CUDA cores. The FP64 cores are sort of useless in the consumer, especially gaming space. The GP100 also featured HBM2 memory running across a 4096-bit interface.
With this super complex and high-end chip and the more simple GP104 that powers the GTX 1080/1070 there is definitely room for a middle-ground chip. There is no way NVIDIA could sell a consumer-based card with the GP100, at least one that would be cost-effective. So we have the GP102. The GP102 will be targeted at high-end consumer graphics and will make up GeForce GTX products including the GTX 1080 Ti.
The chip will be cost-effective as it does away with the FP64 CUDA cores that are found on the GP100. So you have 3840 FP32 CUDA cores, which is 33% higher than that of the GP104. Just like the GM200 had 33% more CUDA cores than the GM204.
GDDR5X is not completely out of the question with this chip either. NVIDIA could make use of the more readily available GDDR5X memory instead of HBM2. We will have to see if the GTX 1080 Ti features all 3840 CUDA cores present on the chip or if some are disabled to improve yields. We also do not know if this chip will be used in the next Titan card or if NVIDIA will go with the GP100.