There will be two versions: the GM107-300 with 768 CUDA cores, 64 TMUs, 16 ROPs and a 128-bit memory bus, and the GM107-400 with 960 cores, 80 TMUs, 16 ROPs, and 128-bit memory bus.
The biggest benefit versus Kepler seems to be a dramatically lower power consumption. The GM107-300 for instance has nearly double the CUDA cores of the GK107-450 while sticking to roughly the same TDP.
Everyone expected Maxwell to be on 20nm process from the very beginning, but TSMC was not ready to manufacture it. Honestly, I’m not even sure if we are going to see 20nm GPUs in Q2. Maybe it’s just too early to talk about them, but one thing is certain though. NVIDIA did not call these cards with GeForce 800 series nametag, because it will be reserved for the real 20nm Maxwell.
It doesn’t look like GM107 is bringing any new technology to the table, it is just better binned and much more efficient processor. Technically you could easily call Maxwell a Kepler Refresh².