Working backwards:
290X =
294W / 2.5 perf/watt =
118W for Polaris 10 ~ 290X (Best case scenario!)
1070 supposedly beats Titan X. Let's just say it = Titan X =
288% (1440p)
Assigned 1440p performance rating:
290X =
207% (118W equivalent Polaris 10)
1070 Projected minimum Titan X level = 288% (150W TDP for 1070)
1070 is ahead of our projected Polaris 10 by 288% / 207% =
39%
Polaris 10 would need a 40% increase in GPU clocks / performance to catch up.
Polaris 10 = 118W x 1.39% =>
New 164W TDP
Major caveats:
1. Assumes increasing performance with Polaris 10 scales 100% linearly with TDP. We know this isn't true for almost any GPU architecture. Firstly, we know that almost no GPU scales 100% linearly with increased clock speeds. Secondly, we know that to increase GPU clocks 40% without increasing voltage or running into exponential increase in power usage is unrealistic.
2. Assumes AMD's 2.5X claim over 290X is 100% correct in almost all cases, not just best case scenarios. If not, the 118W Polaris 10 ~ 290X would mean the power usage of Polaris needs to rise above 118W to match a 290X. For example, if AMD's real world result is 2X perf/watt over 290X, Polaris 10 would be now using 147W to match a 290X. This just shows how much variation there is in AMD's claim of Polaris 10 delivering "up to 2.5X perf/watt" or "2x perf/watt" etc.
That's before we even got to Polaris 10 vs. 1070's overclocking headroom. We are already seeing 2.1Ghz on air from GP104, with rumors of 2.3-2.4Ghz on air and 2.5Ghz with an AIO CLC! AMD's cards in the last 2 generations weren't great overclockers. This once again should give 1070 a big edge.
Since NV claims 1070 > Titan X, and using AMD's own claims of 2.5X perf/watt are most likely too optimistic, Polaris 10 has no chance of beating 1070 in perf/watt using math and neither will it match 1070 FE, nevermind 1070 AIB. More likely than not, NV will also have lower GPU Boost on 1070 than 1080 to create an artificial gap that justifies the 1080's price premium. This means that 1070 should have even bigger overclocking headroom in % terms than 1080 does since they are the same GP104 silicon.
Conclusion:
Using the process of deduction, Polaris 10 will not match 1070 on average even at 150W TDP. It will need 165-175W TDP at least but then even if some AIBs make Polaris 10 cards with max 175W TDP headroom, so will NV's AIB partners who will add 15-20% higher GPU Boost clocks on 1070.
How is my calculation wrong?
If 1070 and Polaris 10 are equal in performance, it means a 314mm2 chip (35% larger) is only as fast as a 232mm2 chip (Polaris 10). This means AMD would have pulled off a 35% increase in performance /mm2 over Polaris 10. I don't believe this is realistic.
> If we use die sizes, Polaris 10 loses
> If we use 2.5X perf/watt claims from AMD, Polaris 10 still loses
There is no realistic math that can get us to the point where a Polaris 10 chip is as fast as a 1070. In all my calculations, I also gave AMD the best possible chance and I low-balled 1070's performance by equating it to Titan X.
I just want to remind you once again that before 1070/1080 came out and we heard that Vega is a late 2016/early 2017 launch, no one really contested that Polaris 10 ~ 390X or Fury at 110-130W. Now that we saw how good 1070/1080 are, and people realized AMD will cede the entire $350+ market segment for 6 months+, suddenly Polaris 10's performance keeps going up and up and up. It's becoming a joke to try and fill in performance gap that AMD has when Polaris 10 is Pitcairn HD7850/7870 successor (clearly), while GP104 with > 300mm2 die size and GDDR5X are straight up GTX670/680 successors.
No one would have expected 7850/7870 to be competing with 670/680 during the start of that generation. So why now are people expecting Polaris 10 -> Pitcairn lineage to suddenly start competing with 1070/1080? It's wishful thinking because AMD is way behind with Vega 11 and 10.
AMD's own comments = Polaris 10 = mainstream + high-end laptops (ONLY)
AMD's own slides show Polaris 10 as a mid-range chip at best
AMD's own comments paraphrased = "We aren't concerned about NV's Maxwell since they are targeting high end"
Just face it, AMD is 6-8 months behind GP104 and they got caught off-guard about how much faster 1080 is + its overclocking headroom vs. what they expected. They also didn't expect 1070 to cost $379, likely expecting it to be $450 minimum.
Even their invite to Computex focuses on APUs, not Polaris 10.
"During this event, AMD executives and special guests will introduce new,
comprehensive details on many of the 2016 products, including the
official launch of the 7th Generation AMD A-Series APU."
This is a big red flag. Sounds like they may reveal details of Polaris 10/11 cards but still no hard launch in June.
http://www.amdcomputex.com.tw/
Not conclusive. Rumour had it that TSMC didn't have much time to optimize A9 for 16FF. Besides, you cannot have it both ways -- if you argue that AMD's chip is more dense, then you need to assign a GPU clock penalty for Polaris 10 over GP104. This is something we've seen in the last 4 years. AMD's denser designs have lower GPU clocks and overclock less than NV's.
AMD's best hope lies in them targetting as many AAA developers as they can and locking in DX12 games/game bundling with Polaris 10.
For example, offer Polaris 10 at $299 but throw in Total War Warhammer with it.