Also, I do believe that 14nm/16nm + new architectures + HBM2 will provide a bigger boost in performance than this generation will. Since HD7970/R9 200 series have depreciated so much at this point, it's not as if selling those cards now is strategic anymore -- that point is long gone.
I will take a bet that perf increase from R9 290X to R9 390X (> 50%) will be greater than from R9 390X to R9 490X (< 40%). Same for GP104 vs GM200 Titan-X compared to Titan-X vs GTX 780 Ti. People are going overboard here with 16/14nm FINFET predictions. This node is the most difficult node ever both for Intel and foundries. Heck even the mighty Intel had serious challenges ramping 14nm FINFET.
I am going to dispute this. AMD and Nvidia have been steadily rising prices for the last few gen.However, if you buy a $600-800 28nm card today such as GM200, that card will likely lose $200-300 in value in 24 months. Upgrade to 2 of those and you are looking to lose $400-600 in resale value because 14nm/16nm GPUs will level this gen's flagships as far as price/performance goes - you can count on it (i.e., recall $330 970 ~ $700 780Ti just 10 months from launch. OUCH). Therefore, if money is a factor, if someone missed the 'perfect' time to resell his/her old card, it isn't such a bad idea to skip a generation(s) if a gamer is OK turning down some settings and doesn't need to run everything on Ultra in every game.
HD 5870 - USD 380
HD 6970 - USD 370
HD 7970 - USD 500
R9 290X - USD 550
R9 390X - ? My guess is atleast USD 650 - 700 as I believe R9 390X will legitimately compete with GM200 for the GPU crown and will be as fast or faster.
GTX 480 - USD 500
GTX 580 - USD 500
GTX 680- USD 500
GTX Titan - USD 1000
GTX 980 - USD 550
GTX Titan-X - USD 1000
The 16/14nm FINFET processes are much more costlier compared to 28nm and with much lower yields. HBM2 will only be ramping in Q2 2016 and that should be costly as well.
http://www.kitguru.net/wp-content/uploads/2015/03/sk_hynix_tsv_roadmap_hbm.png
It will take further time to see HBM2 in end products (lead time from Hynix to AMD/Nvidia) so the earliest I would say is Q3 2016. Remember HBM was in high volume manufacturing in early Q1 2015 and we are going to see the GPU products roughly 5 months later. 2.5D GPU manufacturing lead times are longer than traditional GPU manufacturing lead times.
Further more Nvidia has zero experience with 2.5D GPU manufacturing so expect its fair share of challenges too. And you still believe that 16/14nm FINFET GPUs will level this gen's flagship as far as price peformance goes. Well I would say you are too optimistic. You can expect a perf improvement something like the GTX 580 - > GTX 680 transition. Anything more is unrealistic.
http://www.hardwarecanucks.com/foru...616-nvidia-geforce-gtx-680-2gb-review-29.html
As for prices of the first FINFET flagship GPUs do not be surprised if they move upward to USD 650 - 700. So I would say something similar to say a GTX 580 to GTX 680 improvement in price/perf (roughly 35% better) . I am going to go out on a limb here and say that we should prepare ourselves for a 250 -300 sq mm mid range FINFET GPU with 8 or 16 GB HBM2 at USD 650-700. btw I do not expect the first FINFET flagship GPUs to be more than 40% faster than the R9 390X and Titan-X. I do not expect a > 500 sq mm 16/14nm FINFET FPU before Q3 2017. We will see true 2x leaps wrt R9 390X and Titan-X sometime in 2017 more likely mid-Q3 2017.
I think people are forgetting that Intel has been selling tiny 82 sq mm 14nm FINFET Core M dies for the last 6 months and are only now beginning to to ship > 130 sq mm Broadwell quad core dies in volume. The Intel Xeon chips with die sizes > 300 sqmm are not likely to ship before Q1 2016. If thats how difficult it is for Intel who has 3+ years experience with high volume manufacturing on 22nm FINFET and is on its 2nd gen 14nm FINFET what is the hope for TSMC and Samsung/GF who are only now beginning to ramp their first FINFET products ever and that too mobile SOCs at die sizes < 100 sq mm and <5W TDP.
I repeat the majority of GPU volume for the full year 2016 will be on 28nm and I expect the ratio of 28nm to 16/14nm GPUs from both Nvidia and AMD to be roughly 2:1. I do not expect to see a 16/14nm FINFET GPU before Q3 2016. So for 1st half the 16/14nm GPU volume will be 0. The ramp of 16/14nm GPUs will start in Q3 and pick speed in Q4 2016 and H1 2017.
I also predict AMD will have a smoother transition to 16/14nm FINET GPUs with HBM2 due to their experience with 2.5 manufacturing with R9 390 series and their role in the co-invention of HBM with Hynix. I also expect AMD's 2nd gen HBM memory controller to be better than Nvidia's first gen HBM memory controller. Remember the HD 5870 could run GDDR5 at 1200 Mhz while GTX 480 could run at only 900 Mhz. AMD had experience with HD 4870 which ran gddr5 at 900 Mhz and was thus able to improve on that and run at faster speeds than Nvidia.
In summary do not have unrealistic expectations from FINFET GPUs in 2016. There will be huge leaps in power efficiency but given the yield difficulties I expect AMD/Nvidia to be conservative with die sizes (< 300 sq mm). I also foresee a bump up in prices to USD 650 - 700 especially if there are yield challenges with both FINFET, HBM2 and 2.5D manufacturing. Furthermore AMD/Nvidia are waiting in line after Apple and Qualcomm for FINFET wafer allocation and so the 16/14nm GPU volume in 2016 is going to be much lesser than 28nm GPU volume.
Last edited: