Sure they got a few more out to customers in the first round - but then things were almost as dire for awhile, IMO you gotta give it a couple more weeks to see if nVidia can manage to get stock levels upto scratch or not before making too much of a big deal about it compared to the 5 series launch. If we get into may and nVidia do manage (hah) to ramp up quantities then its not so bad, if come may they are still struggling along at ~10 units a week then its makes the 5 series launch look good.
If they carried on with the same architecture that would be true (and true) but AMD/ATI have let slip several times that Northern Islands is a departure from their previous architecture. From the little info I have I suspect that they mean to split the tessellation unit up into a small number of slightly less powerful units attached to shader clusters similiar to GF100 but not inline with the SMs in the same way and a change to the balance in number of SPs and clockrate.
Again you choose to ignore what I say then add your own spin. The 6 weeks of trouble was because a production line at TSMC was literally not working and had entirely screwed up 6 weeks supply of cores. The supply during that time was similar to what Nvidia is providing now, the launch quantity is some 40-50times higher but you claimed the launches were similar.
Imagine Nvidia losing 6 weeks of wafers due to a screw up, now ADD that problem on top of their current supply because that was the situation a month or so after the AMD launch, even after they seemingly switched out some production of 5770's for 5870's, that still takes 6 weeks to take effect because thats how long a wafer takes to make. After the problem they were back up to normal supply which is 10k's per month, Nvidia has had almost double AMD's lead in to launch time to get a larger supply, yet have had nothing even close.
Again I'll point out you said the numbers were similar, they are not, they aren't comparable, AMD is able to get out 20-30x's what Nvidia can a week at the moment and Nvidia SHOULD have more because they've had several months to stockpile chips for a launch, yet managed to release less than AMD can get out in a couple weeks. To compare them, without laughing or a smilie, is just disingenuous, you're saying 2-3k, and 70-80k, are similar. Compared to the low end, its not a big difference compared to each other, they aren't in the same league and Nvidia's numbers will reduce as supply dries up.
AS for the architecture, you quoted what I said, shader power roughly doubles, I used the number as an example because its easy to understand, you QUOTED ME where I said thats unless shader architecture changes. Complete architecture change WILL NOT change the rough doubling of shader power, even if they went from 1600 current shaders to 12 "new shaders" the rough difference between generations is trying to double shader power.
As for the diminishing returns, again rubbish, every article on GPU architecture will specifically tell you incredibly parallel situations like GPU computations do not hit diminishing returs at these levels.
As for performance, check out newer games, older games that really weren't hard on the architecture didn't show double the power. Metro 2033 on the same settings is TWICE as fast on a 5870 as a 4890. That shows INCREDIBLE scaling, absolutely incredible scaling considering only shader power, rops, tmus were roughly doubled, bandwidth was not and so were several other things. Considering most generations aim to go roughly 80-90% faster than the previous generation, which would be the 4870, not the 4890, which is a heavily tweaked and overclocked version basically, getting double the speed of the 4890 is truly outstanding.
The 5870 shows entirely no diminishing returns and we're years away from seeing diminishing returns.
As for guessing that AMD's tesselation units will change to be more like GF100's, well, I'd say its possible but incredibly unlikely. Why, because AMD have had ridiculous success with small core strategy, spreading out core logic increases die size, not reduces it, I'd put the chance of them spreading out tesselation requiring more core logic and more cross core communication as well, about as unlikely as any real changes that could happen. Keep in mind Fermi is a new architecture to the gt200b, and yet the shaders are incredibly similar, fundamentally the same, a complete new "architecture" yet doubling of shaders that are almost identical. The main reason for the big change from x1900/7900 to 8800/2900 was to be compatible with DX10 and being fully programable by nature, neither is likely to massively change their shader style for a couple DX revisions. The thing likely to change most, is the uncore/core logic stuff, not the shaders.
But I'll point out again, I was using numbers as an example, shader power doubling, across essentially a full node drop, is GOING to happen, it would likely happen with a half node drop to 32nm, they actually have the potential to pretty much triple shader power due to the full node drop and the massive die size saving, however they are more than likely to keep a LOT in reserve because 28nm is due to be with us a long time so it makes more sense to increase performance in steps rather than make everyone pay for a 5000 shader core that no game at all needs for a couple years.
I think its gonna be a couple of years before we see things return to norm.
