Qualifing my "4K is for SLI for now" comment.
Roughly 7.4 Gflops are necessary at 4K to match the PS4's perf/pixel at 1080p. Process scaling looks good for a while longer (even if cost/area increases), so the core issue is the GPU power wall (performance only scales with increase in perf/W). Looking at NVIDIA (since AMD has not released anything talking about an R10 yet), NVIDIA has released graphs of perf/W scaling expectations for Maxwell which place Maxwell at around 2x the perf/W of Kepler. Doubling perf would be 6 Gflops (2x 680) and 9 Glops (2x Titan), which windows the 7.4 Gflop target.
The next logical question is what chip sizes will exist for Maxwell, what processes, and when will they be released? Both AMD and NVIDIA this round released their smaller high-end chips first: AMD HD 7970 at 352 mm2 years before R9 290X at 438 mm2, and NVIDIA GTX 680 at 294 mm2 a year before Titan at 561 mm2. Guessing this makes sense from the standpoint that as the process matures, the yield increases (very important for larger chips). Lets just assume both AMD and NVIDIA might do the same again (a larger chip after the smaller high-end chip), and given the assumption of 2x scaling, that places the single-GPU 4K perf/pixel matching PS4 at 1080p somewhere around the 2014 to 2015 time frame assuming some Maxwell chips ship in 2014 on 20nm (note, VideoCardz.com says that 20nm won't be ready until June 2014).
Other Notes
Looking at the run to 28nm things look good. Flops/area increases, and flops scales with 1/(process_nm^2). But given the number of GPU refreshes on 28nm, seems like process tech is slowing down.
year GPU_______ process mm2 mtrans gflops flop/tran gflop/mm2 gflops 1/process2
2014 R9_290X___ 28_____ 438 6000__ 5000__ 833______ 11.42____ 100%__ 100%______
2012 HD7970_Ghz 28_____ 352 4313__ 4096__ 950______ 11.64____ 82%___ 100%______
2010 HD_6970___ 40_____ 389 2640__ 2703__ 1024_____ 6.95_____ 54%___ 50%_______
2009 HD_4890___ 55_____ 282 959___ 1360__ 1418_____ 4.82_____ 27%___ 25%_______
2013 Titan_____ 28_____ 561 7080__ 4500__ 636______ 8.02_____ 100%__ 100%______
2012 GTX 680___ 28_____ 294 3540__ 3090__ 873______ 10.51____ 69%___ 100%______
2010 GTX 580___ 40_____ 520 3000__ 1581__ 527______ 3.04_____ 35%___ 50%_______
2009 GTX 285___ 55_____ 470 1400__ 1062__ 759______ 2.26_____ 24%___ 25%_______
From my understanding (I'm no expert, I could be wrong here), the wall started at round 20nm. Process scaling is now in the form of moving from a planer structure to a FinFET structure, next bump after 20nm is the switch to Extreme Ultraviolet Lithography which looks like a 20nm to 14nm size adjustment. Collecting internet sources for process dates,
???? ??????????? (marketed as "5nm")
2017 EUV ??????? (marketed as "7nm")
2015 20nm FinFET (marketed as "10nm")
???? 20nm FinFET (marketed as "14nm")
2014 20nm FinFET (marketed as "16nm")
2014 20nm planer
2012 28nm planer
2010 40nm planer
2009 55nm planer
Roughly 7.4 Gflops are necessary at 4K to match the PS4's perf/pixel at 1080p. Process scaling looks good for a while longer (even if cost/area increases), so the core issue is the GPU power wall (performance only scales with increase in perf/W). Looking at NVIDIA (since AMD has not released anything talking about an R10 yet), NVIDIA has released graphs of perf/W scaling expectations for Maxwell which place Maxwell at around 2x the perf/W of Kepler. Doubling perf would be 6 Gflops (2x 680) and 9 Glops (2x Titan), which windows the 7.4 Gflop target.
The next logical question is what chip sizes will exist for Maxwell, what processes, and when will they be released? Both AMD and NVIDIA this round released their smaller high-end chips first: AMD HD 7970 at 352 mm2 years before R9 290X at 438 mm2, and NVIDIA GTX 680 at 294 mm2 a year before Titan at 561 mm2. Guessing this makes sense from the standpoint that as the process matures, the yield increases (very important for larger chips). Lets just assume both AMD and NVIDIA might do the same again (a larger chip after the smaller high-end chip), and given the assumption of 2x scaling, that places the single-GPU 4K perf/pixel matching PS4 at 1080p somewhere around the 2014 to 2015 time frame assuming some Maxwell chips ship in 2014 on 20nm (note, VideoCardz.com says that 20nm won't be ready until June 2014).
Other Notes
Looking at the run to 28nm things look good. Flops/area increases, and flops scales with 1/(process_nm^2). But given the number of GPU refreshes on 28nm, seems like process tech is slowing down.
year GPU_______ process mm2 mtrans gflops flop/tran gflop/mm2 gflops 1/process2
2014 R9_290X___ 28_____ 438 6000__ 5000__ 833______ 11.42____ 100%__ 100%______
2012 HD7970_Ghz 28_____ 352 4313__ 4096__ 950______ 11.64____ 82%___ 100%______
2010 HD_6970___ 40_____ 389 2640__ 2703__ 1024_____ 6.95_____ 54%___ 50%_______
2009 HD_4890___ 55_____ 282 959___ 1360__ 1418_____ 4.82_____ 27%___ 25%_______
2013 Titan_____ 28_____ 561 7080__ 4500__ 636______ 8.02_____ 100%__ 100%______
2012 GTX 680___ 28_____ 294 3540__ 3090__ 873______ 10.51____ 69%___ 100%______
2010 GTX 580___ 40_____ 520 3000__ 1581__ 527______ 3.04_____ 35%___ 50%_______
2009 GTX 285___ 55_____ 470 1400__ 1062__ 759______ 2.26_____ 24%___ 25%_______
From my understanding (I'm no expert, I could be wrong here), the wall started at round 20nm. Process scaling is now in the form of moving from a planer structure to a FinFET structure, next bump after 20nm is the switch to Extreme Ultraviolet Lithography which looks like a 20nm to 14nm size adjustment. Collecting internet sources for process dates,
???? ??????????? (marketed as "5nm")
2017 EUV ??????? (marketed as "7nm")
2015 20nm FinFET (marketed as "10nm")
???? 20nm FinFET (marketed as "14nm")
2014 20nm FinFET (marketed as "16nm")
2014 20nm planer
2012 28nm planer
2010 40nm planer
2009 55nm planer