That's why I said most likely. Certainly all sources do not point to 20nm as guaranteed. Raghu provided a bazillion links over the last 6 months that also support the 28nm theory. I can't tell you 100% if it will be 20nm or 28nm as I do not have insider knowledge. However, neither GloFo nor TSMC have high performance 20nm node. The SoC 20nm may be good enough but I don't recall such a low power node used for SoC being adopted for high end GPUs.
There are some other factors involved. Considering NV goes for the largest dies in the industry, they would benefit from 20nm more than anyone. However, there are no rumours at all that GM200 is 20nm; it's actually the opposite of most sites expecting GM200 on 28nm. One has to ask why would NV with much higher volume sales and much stronger bargaining positioning for prices (as a result of higher sales than AMD) is not likely to adopt 20nm for GM200? That makes it less likely for AMD to do it.
Further, 20nm would have brought a substantial reduction in power usage but yet rumours / leaks say AMD is considering Hybrid water cooling and ChipHell implied 390X would have similar power usage to a 290X. As I said I wouldn't rule out 20nm completely but I think we can't rule out 500-550mm2 massive 28nm die 390X either. Historically ATI/AMD never made 500mm2 die but we have seen them go to 438mm2 with Hawaii. Now 28nm is even more mature, even cheaper and considering high voltage and high clocks contribute a lot to high power usage, going 1Ghz or less on a very wide chip in terms of functional units is better if the costs and yields justify it. In the past ATI/AMD would quickly transition from one node to the next making this strategy too costly for them to execute. Now 28nm would be on its 3rd iteration (7970-->290X --> 390X?).
Additionally, it's very risky to try to pull off a trifecta of new architecture (GCN 2.0) + new memory standard (HBM) + most cutting edge node (20nm) simultaneously. This introduces no room for error. Imo pulling off 2 of those with proven 28nm and hybrid water is less risky.
Finally, there is a matter of volumes, prices and yields. With 20nm so new, and Qualcomm, Samsung, Apple basically outbidding NV/AMD, it'll be very expensive to try and do a top-to-bottom mobile and desktop 20nm roll-out for AMD. They would have needed to start mass manufacturing now or next month to try and get Spring 2015 roll-out in volumes.
Don't forget that there are constant rumours that Qualcomm's 20nm 810 and 615 are overheating. Now if 20nm has trouble with such small SOCs, do you think it's realistic to utilize this process on a 350+mm2 high end GPU?
While AMD did state that they will have 20nm products out in 2015, that could easily be shrunken PS4/XB1 APUs.
As I said, whether or not AMD launches 20nm or 28nm GPUs, it matters more that they hit the necessary performance/watt, absolute performance and price/performance targets. If 28nm allows them to scale GCN another 40-50% beyond 290X, it will be perfectly fine to last them until 2016 when lower nodes become accessible.
Anyway, tying all of this back to Witcher 3:
1) The game could still be delayed 3rd time;
2) Until we know if NV or AMD run this game better, it's a guessing game, somewhat in favour of NV due to the title being GWs;
3) 5 months away is a long time in GPUs; something much faster can be out by May 19.
I wouldn't upgrade right now for the sole purpose of playing TW3. I also would not buy a 4690/4790K at this time. If I had to build a new system now, it would be 5820K or wait until BW.
Last edited: