It looks like NVIDIA will have not one, but two "big chips" based on the "Pascal" architecture. The first one of course is the GP100, which made its debut with the Tesla P100 HPC processor. The GP100 is an expensive chip at the outset, featuring a combination of FP32 (single-precision) and FP64 (double-precision) CUDA cores, running up to 3,840 SPFP and 1,920 DPFP, working out to a gargantuan 5,760 CUDA core count. FP64 CUDA cores are practically useless on the consumer-graphics space, particularly in the hands of gamers. The GP100 also features a swanky 4096-bit HBM2 memory interface, with stacked memory dies sitting on the GPU package, making up an expensive multi-chip module. NVIDIA also doesn't want its product development cycle to be held hostage by HBM2 market availability and yields.

NVIDIA hence thinks there's room for a middle-ground between the super-complex GP100, and the rather simple GP104, if a price-war with AMD should make it impossible to sell a GP100-based SKU at $650-ish. Enter the GP102. This ASIC will be targeted at consumer graphics, making up GeForce GTX products, including the GTX 1080 Ti. It is cost-effective, in that it does away with the FP64 CUDA cores found on the GP100, retaining just a 3,840 FP32 CUDA cores count, 33% higher than that of the GP104, just as the GM200 had 33% more CUDA cores than the GM204.

It could also not be improbable that NVIDIA could use the more readily available GDDR5X memory interface on this chip. It remains to be seen if the GTX 1080 Ti features all 3,840 CUDA cores present on this chip, or if some are disabled to improve yields. It will also be interesting to see on what chip (the GP100 or GP102) the next GTX TITAN SKU will be based on.