TDP is a difficult story, well known to me, but finally i need some measurement spec in order to talk about. In fact it is way more complicated and always hard to compare to each others. If i would truly take any factor into account i could write almost a book, not gonna work in a post and majority of users will run away... simply to much to handle and actually they just want some simplifications.
The efficiency and leakage can be different with every chip, adding another complicated layer to a matter already complicated. People want some "hard specs" but in fact it doesnt exist, every chip is simply different, but still not as different than humans.
I dont have much experience in CPU cooling but you have to separate
amount of heat a chip will produce (i call it TDP but its actually a crazy simplification) and
the amount of heat a cooler is able to "draw" away per mm3. The experience i had with Intel CPUs, in my case 3570K, is that it is very difficult to "draw" away the heat from the IHS, even if the TDP is actually very low. So this low 77W TDP chip is actually demanding on a cooler. I had a easyer time cooling down a 990X with comparable cooler because the heat is arriving at the sink, it does transfer well,... this is a 130 W TDP CPU in theory.
@Ivy, I'm not sure if you have any idea about Maxwell scalling with voltage. You can set 1500MHz+ without raising voltage on most GTX970 and many GTX980 what barely raises wattage. GTX750/Ti are scalling in similar way.
AMD are pathetic if you look at efficiency scalling while overclocking. Look at every series in last 5 years. 7970 needed huge bump in voltage to make 200MHz more. 290/290X the same so without really big changes in architecture ( right now there is not much except that memory bandwidth thingy ) I don't expect anything great from new series.
Maxwell is actually a sly little wolf in sheeps clothing, because most cards can not be Volt "bumped" without some mods or special software. However, it does work differently, it is simply capped by thermal and power design and as long as thermal and power is not upped, any other matter is secondary. I have graphs showing me that the Maxwell is losing a lot of efficiency at high clocks (from 100% down to ~65% at above 1500 MHz), even without Volt bump, so
a Volt bump is not a requirement in order to lose efficiency. I bet many 970/980 are far above 200 TDP without any Volt bump, yet people barely notice it and are telling us "that card is doing the clocks out of nothing..." but this is not true. Nvidia simply used a dynamic approach because they know pretty well that the difference in thermal and power capacity can vary a lot and not every PC or card is able to handle it, so a very dynamic design is the result of those facts.
I would imagine that a reference card with water cooling would benefit from a mild overclock- let's say it uses an extra 25 watts from a stock 225w tdp at 1000mhz core clock and increases speed to 1250mhz. This could make it as fast as a GTX 980. But let's say there's another vendor that wants to overclock it to 1500mhz, way more than the reference design. Having the water cooler helps and is what might make the TDP reach around 300 watts. That would appear to be normal by today's comparisons, but it shows that the fastest cards don't have to start at super high TDPs just because they're expected to...
I actually have a hard time following you, more than most of the other stuff. The issue is, yes AMD was already adding reference water coolers on high end CPUs as far as i remember, this approach was already done in the past. But it is a totally different situation, AMD was already using the whole TDP headroom their CPU was able to get, and the only chance trying to barely compete with Intel was to add a water cooler and selling it as a "water cooled CPU", making some price-sensitive enthusiasts happy and trying to increase reputation on the "benchmark list" this was the whole reason for doing it, not the very low... truly very low margin.
The situation on the Radeon cards is different. They dont lack the power in order to compete, they have a realistic chance to compete without special solutions. The only risk is the
flagship single GPU because this GPU is headed toward enthusiasts. It is a minor share on the market but it can be critical for marketing and reputation purposes: "hey they got a formula 1 car, we are gonna buy from this formula 1 car manufacturer, but just a street car still". Finally its not just in order to serve the few enthusiasts, but to set some "mark" in the mind of any customer... even to those that dont buy flagship stuff. So the thing AMD may want to achieve is to drive the flagship single GPU to insane levels
out of the box, not after some "do it at your own risk you sorry user" approach. Now, a crazy "out of the box OC" is not gonna do the trick, because it will become bad at efficiency, bad at possible headroom and finally bad reputation, in that term the target was failed. AMD will have to use a sheer ressource pumped GPU, kinda comparable to the 290X in order to achieve the target. The 7970 was already bumped a lot at release, AMD never made such "ressource bumps" in the past, but theyr target is clearly to beat... to impress and to set a mark. The hunt for margin was not high releasing those ressource bumped stuff, but it did pay off for them because the 7970 for example was a very successful card, so the risk of the investment was actually fully covered. I guess they will continue doing so... but for flagship only because it will set the important "psychological mark" in the mind of any user, not just enthusiast, so it may be pretty possible that the next single core flagship can become a 300W water cooled reference card... who knows. There isnt a huge market share for users getting it but thats not the point, they wont have a wield high enough for mass production anyway. The big money is done with mid range cards... not with flagship cards, they have another role. AMD clearly want to prevent the situation users telling in forums "oh... i am gonna use a SLI 980, better than the 390X", it would destroy a important psychological mark, so target is to offer something competitive even for the "i love SLI" users.
Regarding 225W TPD cards... no they will not get a water cooler, it is against any economical rule nor target. Those cards are kinda same such as a 970, a good gamer card at affordable price. Water cooler may be used on 300W+ and no less. In order for OC, there will be many aftermarket designs from third party... it was always like that and it will stay this way, at least for any card not "flagship grade" and actually only enthusiasts may enjoy such stuff, not a big market but a market that got a voice on the web.
So, what is very much possible, is if this pattern holds true for two generations, we'll see a "top of the line" R9 390x around 200-225 watts or possibly 250 watts (possibly on a new nm), and then we'll see the year after an r9 490x with a TDP of 275watts or 290 watts again. It's not that the GPU can't handle more TDP, it's just historically imitating the tick-tock strategy of Intel where it uses a smaller die, then "maturing" the generation after, with more shaders, etc
Many interesting thought but i think it is a pretty feisty matter for the 300 series going your way because the 300 series is a "tock" (new architecture) already, this is almost certain because
AMD can not beat or match Nvidia using a "tick" (20 nm shrink without new architecture) regarding the 300 series, it would be impossible reaching the required efficiency and performance even if they gonna use a 300 W TDP design. So this option is unrealistic and because of a still immature 20 nm node this matter would end into a financial fiasco, i do strongly disapprove doing so. So, taking this matter into account it cant be a tick (20 nm node shrink) according to this "rule". On top of that, AMD didnt release a new architecture for over 3 years already (it was all based on the old GCN with only minor improvements at some spots). So there is some reputation to hold up and in order to do it, the best deal is to make serious investment and not looking at the margin to much which is historically not a weak spot from AMD "looking at the margin". Nvidia is the one trying to make a margin battle and surely very successful doing so. AMD since the release of GCN was not going the "margin battle way" but they was on a road in order to truly set some reputation mark in many terms and they surely was successful doing so because their market share in overall is the highest they ever had, so the target is achieved. You have to take into account, the current nm-technology is already on a very high maturity and the chance is high that the 300 series will still be using 28 nm. So this means that the manufacturing (wield of 28 nm even at insane transistor levels) is not a issue and the resource can be boosted a lot (kinda comparable to 290X). The next series, the 400 series, will probably be the same architecture but it will be a "tick"-matter (shrink of the old architecture down to 20 nm), so AMD will probably use the new 20 nm node and as a result increasing resources even more and they can have a large headroom in term a liquid cooler is used as a "reference design". So it makes good sense putting the 300 series at a tock (new architecture), which is almost certain, there is no doubt about because the old GCN is now at the limit of the possible headroom since the release of the 290X which is basically setting the last type of "old gen" chip.
Your ideas could be correct in term AMD is using a "tick" (20 mn shrink) for the 300 series, but i think it is rather unlikely gonna happen. The new 20 nm node is still challenging at manufacturing efficiency, this means AMD would lose the chance to impress by using a "high boosted tock" (new architecture at high ressources on the old 28 nm node). Which could be the better option, and in that term saving up the "tick" (20 nm shrink) for the 400 series at a time of improved 20 nm maturity. The 20 nm, next year or so, would be able to give new TDP headroom, so that means they cold make a "tick" with a ressource boost included by using the same old cooler (maybe liquid cooler, who knows). But it surely would be insane trying to challenge Nvidias Maxwell by using a "20 nm GCN of the old architecture (tick)", it may grant up to 30% more performance at same TDP and it would be insufficient in order to beat a Maxwell.
Especially because Nvidia could then move on using a 20 nm shrink on the Maxwell too and boosting the ressources to twice as much, this means Nvidia in term of efficiency would be so far ahead of AMD Radeon that the Radeon users will look like "weak volcano riders"... a pretty hot issue for sure. The 20 nm Maxwell would be more efficient and more powerful performance both at the same time... so it will simply be foolish trying to ride the old GCN without a "tock" (new architecture). So in my mind, once again, shrink (tick) is not a priority for now... architecture is the key.