That's not very difficult when nVidia hasn't launched yet.
Yes, it's hard to win when you don't show up, but nobody is holding up the race for them. Every day they fail to deliver increases the advantage for AMD. It not only increases AMD's market share and mind share, but increases the chances of them being able to spoil Nvidia's launch with a refresh or review edition which increases pressure on Nvidia to outperform. If they are working up to a part that bests the 7970 by 25% but wind up going against a 7980 at 1.2Ghz (binned or respun) they could wind up having to compete on price rather than performance.
I'm not saying it's dire or anything at this point, but if we're talking past April, Kepler best be impressive.
It's a gamble like in a 100 meter sprint race. You can try and launch a little early hoping that it's not too early or you can play safe and risk giving the other guys a headstart. As I said, in my book it looks like taking the risk has paid off for AMD.
From their business point of view, Nvidia might as well feel like they are on the winning side also, provided that Kepler really is that good in terms of perf/w and they've taken their sweet time in order to secure a lot of mobile market share which might carry better margins for them.
Just saying it's rather ridiculous to talk about AMD doing better than Kepler when Kepler isn't even out yet. AMD obviously has some advantage, but we'll see what that translates to down the road.
You're still holding onto this notion that whoever launches second has failed - amusing considering there are only two companies. There are just as many advantages to "losing" especially if you have a compelling product, HD4870 ring a bell?
There is no long-term first mover advantage by launching a few months early in a mature market and picking up early adopter sales, especially at Tahiti prices. I don't get why people get so worked up over such short periods of time. Can you name another market where launching a new product 3 months after the competition is a failure and giving them a free pass?
Yes, GF100 was late and nVidia was fine. Hence proving my point. AMD may not have provided a public date but it's pretty clear they missed their own targets simply due to the problems they had (see the same Anand article you quoted).
The only thing moving is your interpretation of what I'm saying
If I were to follow your logic then a hypothetical March launch for Tahiti instead of January would mean Kepler is only one month late instead of three. Or even better, if both companies screw up then nobody is late and the customer doesn't mind one bit. Sorry, but that doesn't click.As I said in another Forum, imagine nvidia released this 7970 as a succesor to the gtx 580 (500 Bucks;new High End). People would call it a joke.
Jawed
Legend
So if NVidia's next high end chip is 10-20% faster than HD 7970 that's a joke too? Is that right?
Yes.
I didn't raise the specter of G80 or R600. That was Arty....good try though
Besides, R600 didn't flop due to delays. It flopped due to how it performed in the end. Like Anand said, if AMD had turned the tables on G80 after all the delays things would've been much different. Again I'm not one for digging up ancient history though.
Lol is that a serious question? How in the world is Tahiti vs GF110 in any way relevant to Kepler vs Tahiti? Or are you just taking the piss?
If they do it at a hundred watts less [strike]than 7970[/strike], nobody would.
Last edited by a moderator:
boxleitnerb
Regular
But then it wouldn't be the successor to the 580. With 100W headroom they could make a card that completely smoked the 7970.
"Can" and "Will do" is 2 completely different things, nothing prevented AMD from releasing 100W higher (real load, not theoretical TDP etc) chip using nV's formula for Fermi TDP, and it would be leaps and bounds faster than the current 7970 is.
And if the 7970 had the 580's power draw it would smoke the 580 by an even larger margin than it already does.
Also, the 5870 was only ~40% faster than the 4890. I'm sure nobody here is going to suggest the 5870 was a joke...
Last edited by a moderator:
fixed.
That's the nice about being first. You're eligible to be compared to the last gen, not the first try of the current gen.
It's my understanding (and I'd like to stand corrected by someone with real knowledge on the subject) that hotclocking ALUs might save you some die area but it shouldn't be for free either. In order to answer that question one would have to know what the gain for the hotclock accounted for more or less.
The Kepler/Tesla claim for perf/W is for sustainable DP performance compared to Fermi/Tesla. They've either improved some parts that affect data management in order to save bandwidth or they've simply gone brute force with the top dog to a 512bit bus. Both sounds a wee bit over the top and I don't believe in miracles that easy either.
I think the TeslaM2090 from its 665 GFLOPs DP on paper achieves in DGEMM around 350 GFLOPs DP if memory serves well. Either I'm missing something essential here, or the 2.5x times claim is an understatement if we're talking about sustainable DP FLOPs for GK110.
Other than that I'd dare to estimate that ultimate goal might have been DP FLOPs for the change in architecture; if GK104 truly has give or take 3 TFLOPs for SP, then it obviously won't mean by as much more 3D performance compared to GF114, especially encounting that FLOPs != FLOPs between different architectures. Even worse it's at least twice the FLOPs than on GF110; I'd say the GK104 will be extremely well positioned if it can surpass as a performance chip by a relatively small and reasonable amount a GTX580.
boxleitnerb
Regular
The 5870 had several things going for it. The 8xAA weakness of Nvidia, DX11 first and SGSSAA. I wouldn't call the 5870 a joke, but in part it only looked so good because Fermi was late and inefficient. In absolute terms, the 5870 was "normal" for a shrink.
Today, the situation is quite different.
As for the 100W less...the difference between the 7970 and the 580 is only around 40W. Not really much. 100W less would make this hypothetical Kepler card way more efficient than the 7970. Even if both cards rose to the same power consumption (real, not TDP, furmark etc.) level, the 7970 would still be behind by a seizable margin.
Last edited by a moderator:
With half nodes gone, the race dynamics have changed. The lifetime of a particular chip will increase, as we've already seen last time. And there is less improvement possible if you stay in the same process.CarstenS said:
So if Kepler is a decent architecture, the delay is not nearly as bad as it would have been in the past.
(deleted. unnecessary noise. )
)
