IF YOU WERE wondering why Nvidia put out the GTX260-216 with a stupid-sounding name instead of the saner 270 moniker, here is your answer. There is a 270 coming, it will have a big brother called the 290, and a dual card code named "China Syndrome"(1).
Yeah, NV is in deep doo-doo right now. The card that was meant to power its way to profits, the GTX280, went from $649 at launch to $499 a few weeks later. A quarter or so on, it is selling retail for sub-$400 prices here and there. AIBs tell us that the 260 costs an ironic $260 to make, which closely matches the teardown numbers we have seen. Toss in the mandatory 15 per cent markup at the retail level, and if you see one for sale at under $300, someone is eating money. Basically, if you can make money on these parts, something is wrong.
On the up side, the 280 is the single fastest GPU on the market. On the down side people don't buy GPUs, they buy graphics cards, and the 280 is not the single fastest graphics card on the market. That honour goes to the ATI 4870X2 by a large margin. With the new-gen GT200 parts, Nvidia loses on all fronts, performance, performance per dollar, and performance per watt, they simply aren't competitive.
That brings us to the new parts, the 270 and 290. They popped up on a PNY price list a few weeks ago, and then were pulled off immediately. This part is what we were calling the GT200b in May, but the public code name is GT206. It is simply an optically shrunk GT200, so clock for clock, you won't get any speed boost out of it. It is meant to fatten up the margins by reducing cost. If the GT200 is a 576mm^2 die, and the 206 is around 460mm^2 (~21mm*21mm die), even with the more expensive 55nm process, NV should save some money.
One big problem though is yield. GT200s started out at 40 per cent yield, that is for the 280 and the yield salvage 260, and were up to a hair above 60 per cent last time we looked. Toss in the smoothly-named GTX260-216, and you screw up the binning a lot. When you transition to a new process, yield almost always goes down, so this part should be back in at the bottom of the yield toilet, not that 60 per cent is out of the bowl, in short order.
So what are the 270 and 290? That is easy, they are 55nm GT200s, aka GT200b, aka GT206. Nothing new, nothing spectacular at all. Why the new name then, other than desperation, if you get nothing different clock for clock? Well, that is easy, Nvidia simply has to bump up margins, and the easiest way to do that is to snow customers.
If you remember what they did when changing the name on the 8800GT to the 9800GTX, it is the same thing. Partners can't make money, and a name change will make the stupid fanbois out there line up. Milking the stupid is a time-honored tradition in the GPU world, and since the 55nm shrink of the G92, the G92b sold less worse when they renamed it the 9800GTX+, it looks like they are trying it again. No, not the GT1xx, this time it is going to be less egregiously renamed to the 270 and 290.
When you shrink a chip, there are three main benefits, area, power and speed, with the last two being a tradeoff of sorts. Going from 65nm to 55nm, you shrink to about 70 per cent of the area, and save a bit of power. The power savings, however, are not quite 30 per cent off the top and, in addition to that, you also lose more on leakage, as smaller structures leak more. People who have done the same shrink at TSMC tell us that you net far less than 20 per cent. Let's be generous and call the power savings 15 per cent.
The 280 has a TDP of 236W and the 260 is at 182W, with the 260-216 somewhere in between. Since the RAM, external circuitry, power regulators and fans don't scale at all, we will ballpark the board level power savings at 10 per cent. That would put theoretical 270s at 164W and the 290 at 212W, at the same clocks, but lets be overly generous and call it 160 and 200, just to give NV the benefit of the doubt.
As always with transistors, you can take the power savings as such while keeping the clocks the same, or keep power consumption the same and jack up speed. Guess which one NV is going to do? Hint: "20W less power consumption!" on the side of the box won't sell many cards. You would be better off airbrushing bigger nipples on the chrome chick cover art.
So, in the end, the 270 and 290 are simply a little faster 260s and 280s. Whoopty-fricking-ding-dong. They will be relaunched as the second coming, and sites that are far to afraid to be cut off will parrot back the NV sermon. It won't be fast enough to beat a 4870X2, not even close, but it will let NV jack up prices just in time for Christmas, giving their partners a desperately-needed few points of margin. NV fanbois will line up for it, and wonder why they took such a bath on the $650 280s they bought.
These cards were supposed to be out in late August or early September, but are now set for November. That is about time enough for a single respin, so something must have gone a bit pear shaped. In any case, it is much less delayed than the GT212, but that is another article entirely. AIBs just got their 270/290 boards recently, so both are a few weeks out yet.
If you are underwhelmed, then the dualie card is for you. We haven't seen a code name for it officially, but NV AIBs are talking about it. Take a 55nm GT200b/206 and put two PCBs together a la the 9800GX2, and you get the idea. There is one minor problem this time... heat.
The G92 (8800GT/9800GTX/GT15x) was coolable, barely, with a single slot cooler. The GT200 is not. Even with a theoretical 20 per cent lower power draw, you would be at 290W for a dual 55nm 260 clone. If you use a 260-216 or jack the clock up, you are at 300+W in an instant, and we can see 350W without trying hard.
Normally, you would do what ATI did with the 3870X2 and downclock it a bit. Not much, just a little, and take the power savings. The problem is that the 260 loses already to the cheaper 4870, and two of them wouldn't be much of a fight against the 4870X2. At a minimum, you need 2 260-216s, or better yet two upclocked 260-216s to pip the 4870X2 and claim a hollow victory.
NV is in a real bind here, it needs a halo, but the parts won't let it do it. If they jack up power to give them the performance they need, they can't power it. Then there is the added complication of how the heck do you cool the damn thing. With a dual PCB, you have less than one slot to cool something that runs hot with a two slot cooler. In engineering terms, this is what you call a mess.
Given NV's problems of late with cooling, (here, here, and here) it is in a bind, but there is no way out of this, none at all. The only thing it can do is resort to unethical tactics, and that is what we think it will do here. The only solution open to NV is to cherry pick ultra low power GT200b parts and make a small run of GX2s that don't burn a hole in the bottom of your case on their way down to the center of the earth.
If this case follows past tactics, NV will make a very small run of parts and claim there is full production. Think 1000 parts or so, most of which go to reviewers. The rest will go to high-profile marketers, think Newegg, and they will sell out. When people cry for more, the usual 'high demand' lines will be spun, and they will dribble out 10-20 cards here and there to keep up appearances.
Pricing will likely be right on top of the 4870X2, maybe $50 more in order to bolster the performance claims, and they will be undoubtedly be sold at a loss. Then again, with the number likely to be made, it is chump change to take a small bath on each one to claim the lead for Christmas.
In any case, if the three upcoming parts, the 270, 290 and GX2 look like naked desperation attempts to grab at a halo, you are right. Upping the clocks will hit NV in the bottom line at the end of the quarter, their margins will suffer because of this stupidity. Fanbois will love it, massively subsidised parts are great deals for consumers. In fact, you will likely see massively discounted 260s and 280s in a few weeks when the new parts hit the street. They won't save Nvidia's bacon, but they may help partner margins. For a short time. At a high price. µ
(1) It was code named "Smoking Sepuku", but Nvidia didn't want to publicly suggest that this strategy was vaguely related to anything honorable, so it was renamed. We saw the memo, it was poignant. (2)
(2) We are, of course, making this all up.
Originally Posted by Shintai
slightly lower power consumption and/or higher clocks 
probably has to be less than 600MHz per core due to cooling limits 
Bookmarks