I smell PR bullshit,as much as I like NVidia , they're always talking PR and being cocky,  but I'll just wait and see.

if this is true , we might see huge advancement very soon, 

not soon lol, they didn't even put a date on it so at best we are looking at 2016, with a good chance of that being late 2016 and a good chance of that slipping into 2017. Maxwell is 2014 after all and then we will get a refresh of that before they even consider launching Volta.

How does that projected number, 40TFlop, correspond to the manufactoring process of 10 nm? I have forgotten the maths behind die shrinks (the correlation between process shrink to amount of transistors and power consumption). Could a shrink from 28nm (today) to 10 nm really allow an increase of around 20 x the processing power that we today get from a 100W GPU?

it's their projected goal, we'll find out how close they'll be when the time comes.

So in 2017 we will see a bandwidth of 1 TB/s... seems cool... I thing that is another proof Sony choose GDDR5 instead DDR3 was in the right way... the games in the future will use more bandwidth for sure.

Nvidia are so good at the hyperbole marketing.

Intel beat them to stacked DRAM anyway.

lol? Titans, as in the $999 GPUs? And two of them?

Its roughly 8-9 times more transistors from 28nm to 10nm. So you could get a 8-9 times increase in raw gpu performance just by putting 8-9 28nm Gpus on the wafer, theoretically. But there is no math for that anymore the development isn't as straightforward as it once was.

It also depends on architecture. There are many transistors who have not really something to do with flop performance this comes from the cores. And I doubt Nvidia needs 9 times more of everything. So they probably target to put a higher percentage of transistors to use for flop performance than in current Gpus. Its possible to have a 20 times increase tflop performance from 28nm to 10nm it all depends on the architecture the wafer quality etc. Could be anywhere really without inside knowledge its impossible to say. Nvidia always reached roughly their goals. So a 30-40 Tflop 100-120 Watt Gpu from 2019-2022 can be expected.        

Theoretically that's the deal but I doubt it's going to work.

Since the revolutionary design of the GT 8000 series to the GTX 600 miniaturization went to shrink transistors tenfold but on raw floating point operations we had around a six times increase - of course you could argue Nvidia was focused on delivering  improvements to other areas deemed more important such as double-point operation but die sizes and power consumption increased too anyways and it doesn't offset or excuse it completely. And now from 28nm to 10nm you shrink a transistor around eight times. Take of that as you will... if Nvidia dramatically change their architecture after realizing more raw number crunching could work better it could happen, but following trends or their current design, it won't. 

Edit - and it seems like you have a multitude of opinions to draw your conclusions from. I may be the most pessimistic one but right now looks like to me the mighty fundations of predictable increases seem to be finally faltering as we get stuck for longer and longer on the same manufacturing processes. 

I think Nvidia has to focus more on raw numbercrunching if they want to build Exascale Supercomputers someday. They need to step up their game imo.