I am sourcing this from another forum but it should provide some interesting discussion. I can see that nvidia is going to honestly need to redesign its next node or self there will be many issues with refreshing. It seems from what I have gathered here that simply moving to a smaller node is more challenging with graphics cards then with CPUs.
I can see price seeing a bit of a hike on the cards. After all nvidia has not made very much profit with Volta at all and to skip it after researching it is quite a fiscal hit
Ampere is fix, Q2 2018 too. The first already start with the BoM.
Source: Igor Wallossek of Tom’s Hardware @ 3DCenter Forum
Obviously, Igor once again presented a new internal roadmap, in which now no longer “Gen. 4”, but just “Ampere” is - and now firmly for the second quarter of 2018. This means a significant transformation of nVidia’s plans, in which the Volta-based gaming chips GV102, GV104 and GV106 (as well as possibly others) were certainly once firmly planned - and will now be dropped in favor of the next generation. What this is, is unclear - conceivable would still be the one that has misjudged the manufacturing technology and the actually intended for Volta 10nm production is just too late for large graphics chips ready for the competition. In the case of the HPC chip GV100 , which was definitely to be brought in , it was necessary to resort to 12nm production, which, however, produced an unusually large chip even with 815mm² chip area. This extreme chip area may well be regarded as a certain indication that nVidia had actually planned the GV100 in 10nm production - because in this case the same chip would probably be ~ 500-550mm². And possibly the GV100 chip had to be slimmed down in its final form because of the limitations of the 12nm production even something, the non-fulfillment of some early performance promises points in this direction.For the gaming chips in the 12nm production, the pressure is lower, because you (mostly) does not go into such areal border areas. But for a higher performance under the 12nm production always means a larger chip area, since the 12nm production in this regard has only small advantages over the 16nm production (according to TSMC 20% less area requirement at 10% Mehrtakt or 25% less power consumption compared to 16FF) . Either nVidia would be able to offer gaming chips in 12nm production so rather only below average performance gains - this could be realized with slightly larger chip area and about the same power consumption quite. Or nVidia would have to make gaming chips in the 12nm production significantly larger, which would then at least have a medium-high performance boost result - but this at a higher cost and above all to a higher power consumption. How such a thing can look like was already exhaustively explained in a previous speculation article .
A performance doubling or even the proximity of it is not achievable with the 12nm production - for this the graphics chips would have to be too big and too expensive, especially the GV102 / GA102 would then encounter production limits (the GV100 chip with its 815mm² could be made at all only by means of expensive special procedures) . But above all, an additional performance under the 12nm production is always accompanied by an additional power requirement - really fast 12nm chips would only be possible by doing away with nVidia’s first-class energy efficiency and (relatively) low power consumption. This should probably have been the sticking point, why nVidia has taken distance from this path then again: With the 12nm production for gaming chips, there is either only a small increase in performance, which would not have arrived in general - or nVidia would have to buy a large increase in performance with significant additional costs and the renunciation of the known energy efficiency, which would have been criticized as well.
Either way, the interim solution of 12nm production is not good to adapt to new graphics chips, was ultimately probably only for the GV100 chip a necessary evil to fulfill their own HPC plans or obligations in this regard. For new nVidia graphics chips after the Pascal generation you actually always needed a new real production fullnode - with everything else either the performance or the power consumption requirements would not be met. In addition, the lateness of a new nVidia graphics chip generation (the 18-month hitherto, and probably a total of ~ 24 months after the release of the current generation is comparatively long for nVidia) suggests that you are waiting for something to happen At a certain point in time is ripe - like the 10nm production for large graphics chips. For smartphone SoCs that is already in mass production since spring 2017, but usually the first months are blocked exclusively for large orders from Apple and Samung and subsequently the new production must first mature in such a way that produces the much larger graphics chips to a meaningful production yield can be. One year later than the first corresponding SoCs here is a rule of thumb, which has worked well in recent years - and now in the case of the 10nm production of TSMC fits well with the second quarter of 2018.
Long story short, due to these circumstances, we expect the ampere generation already in the 10nm production of TSMC. After this (for nVidia) long waiting phase between two generation of graphics cards as well as due to the availability of this production method, which was already predicted earlier in the spring / summer of 2018, everything else would really come as a surprise. Above all, the use of 12nm production in the spring / summer of 2018 would hardly make sense, as long as the 10nm production would soon be ready to go - nVidia (as explained above) can not achieve as much performance with 12nm production as you really wanted to bring. With the ampere generation in 10nm production, however, the generally usual performance boost of almost twice is possible. How exactly nVidia achieves this with the (supposed) HPC chip GA100 as well as the (probable) gaming chips GA102, GA104, GA106, GA107 & GA108, is still unknown and is therefore in the field of speculation . Conceivable for this is from larger architectural changes over pure more hardware units up to (still) higher clock rates still everything.