You're thinking in terms of perf/mm. The "sweet spot" is about addressing a market segment - a high end chip can only be profitable whilst addressing one segment; RV770 addressed, more or less, 3 segments within 6 weeks of initial launch.
Q&A with AMD’s Rick Bergman on the graphics sweet spot
- Thread starter Jawed
- Start date
Mintmaster
Veteran
True, but I really doubt the X2 cards are as profitable as a single high-end chip would be, particularly due to the 2GB of GDDR5 and doubled substrates. Designing a separate chip for the high end would have been smarter in a lot of our eyes, and I think executives at ATI would agree if they knew from the start how awesome RV770 would be and how GT200 would perform. Of course they didn't know, so I'm not blaming them for the decision.
Shaidar, I was talking about the high end only, but it's entirely possible that GDDR5 will start to become prevalent in the lower end as well as prices drop. RV740 is $80, so even if GDDR5 pushes it up to $90, I think having twice the rops and BW with all the tiny 40nm SPs that a 128-bit RV840 (or whatever it's called) will pack will be very worthwhile.
And yes, I know that nobody's cracked that egg yet. I'm hoping that ATI can.
RV770's architecture is so awesome that ATI can probably devote 50% of their top architects to that problem and just make evolutionary changes to everything else.
Last edited by a moderator:
I'm not convinced that IHVs cannot make a quite safe estimate what one or even two refreshes of base architecture X might turn out to be. Ok estimates might fall off by a small persentage, but not to the point that the end result would have a huge difference compared to the estimate.
In my mind the whole change of strategy sounds rather like long term planning within the reasoning to better start early than too late.
I don't know frankly what NV did in between, but there are tons of rumours floating around. One speaks of them expecting RV670 to be a 3 cluster design, the other that there was a mystery GT200 with 384SPs that was canned because the rest of the chip was delivering even less performance then today's result etc etc. To me personally it doesn't matter much what the real reasons for each result really are. The real point is that AMD has managed to recapture lost ground from the R600 days and as ATI learned from its past mistakes I believe that also NVIDIA will learn from its own.
In the end the glorious news is that the end user doesn't have to pay nearly 600 bucks nowadays for a high end sollution.
One or two successes on the same reasoning, do not guarantee a third or more successes, unless the competition is dumb enough to not change anything. Frankly though I don't expect NV to be able to make any major difference until their true next generation architecture and even then we don't know squad yet what they've been working on for the past years.
I personally have way too many important question marks regarding anything that might come on 40nm in 2009. I have no idea if AMD's next RV8x0 is going to be D3D11 compliant or not and much more important how much die area the added necessary features would occupy. I'm fairly sure that NV's GT21x@40nm line hasn't anything to do with D3D11, but that's not good enough to make any safe predictions yet.
If however RV8x0 follows a similar concept as its predecessors and is a D3D11 chip, it will give them most likely something like half a year headstart with 11 compliance but I have severe doubts that NV is going to base their own 11 solution on something like the GT2x0 idea. Depending how evolutionary or revolutionary NV's GT3xx architecture will be the tables might remain the same or even turn.
On my personal wishlist I'd prefer high end GPUs to remain at the same price margins as today. If you deduct the price difference between a GTX280 and 260 or on the other hand between a 4870X2 and a 4870, is the performance gap (on a real average on a long game list) really analogue to the first?
hoom
Veteran
Less profitable in hardware cost per card at the top end but has to be much more profitable in terms of being able to have one engineering team working on making one chip (better performance/mm) instead of splitting the talent/resource/funding across 2+ chips as well as $$$ saved on tapeout, qualification etc.
Two things there - first consider the TAM for > $400 graphics cards then also consider the TAM's for <$200 and <$300 boards. Additionally, the memory sizing is a board implementation detail and, although there will be some projections at design phase, board level decisions are made much later in the process. We're talking about the decision process a few years ago now.
Mintmaster
Veteran
What's TAM? Total addressable market?
Anyway, I'm not saying that ATI should skip the $200 and $300 markets! I think it makes sense to bring out the performance parts first, but they completely ignored the high end aside from the X2. 2.5 years ago did ATI know that RV770 ALU size would be so small? That doubling Z-rate and fixing AA would have so little impact on ROP size and improve perf so much? That GT200 would have been weak enough to beat handily with a 350 mm2 GPU?
I think that if RV770's designers knew these things, they definately wouldn't have skipped the high end, even if they did decide that the primary focus should be on the $200-$300 market with that chip coming out first.
I'm sure they had a hunch, at least. Did you read Anandtech's piece about this?
These were purposeful design decisions made to keep die size down. ATi didn't want (or couldn't afford) to make huge monolithic dies anymore.
Silent_Buddha
Legend
Basically it's just a case of minimizing your losses should something go wrong, and at the same time focusing all your efforts at the largest market.
Consider the scenarios where your "flagship" (First to market) card is a large monolithic targetted at the relatively small enthusiast market and the case where your card is relatively small and targetted at the larger mainstream/performance mainstream market.
In both cases you'll have a relatively long developement cycle, that's unavoidable.
In the first case however you have a very large chip that's expensive to make, more prone to defects in manufacturing, and more prone to errors due to the sheer size and complexity.
In the second case, all those factors are mitigated to an extent. Less expensive, less prone to defects per die, less transistors to have a possibility of screwing up.
If you fail in the first case it's an extremely expensive failure with a low possibility of recouping R&D costs. Look at R600 as a prime example.
If you fail in the second case, it's less costly and you have a bit more wriggle room to recoup lost R&D through a larger market and an already low price point.
In the first case, success at best brings you a halo effect for the rest of your lineup and a boost to company image. Profits/revenue still won't be as good as the larger mainstream lineup. However, if the halo effect/image can boost your lineup enough that's moot. Problem is, in many cases the mainstream parts have tended to lag the halo parts by quite a few months the past few years.
In the second case, success hits the sweet spot of the market where the most money is made. It can also arguably cause a halo effect of it's own as you have the possibility of reaping numerous price/performance awards. Company image is further boosted in this important market segment as you earn word of mouth of providing best bang for the buck. And market can always spin that the company with the faster halo enthusiast part doesn't care about the masses and only caters to the elite rich boys, yada yada yada.
In other words. There's a lot of downside to targetting the enthusiast market as your main target and then moving down from there. And not a lot of immediate upside. Even if you execute well consistently.
Focusing on the mainstream/performance mainstream doesn have a whole lot of downside and while the upside isn't nearly as flashy it can be more profitable. As long as you don't forget about the PERFORMANCE. Making a mainstream part and then not having leading or competitive performance will kill you. Just look at S3 which is still trying to get that right.
Unless something drastic happens. I don't see ATI moving away from trying to hit the affordable performance segment in favor of the all or nothing enthusiast segment anytime soon.
And as long as sites benchmark as they have in the past. I don't see them moving away from multi-chip cards for the enthusiast market (even with icky AFR) as it still scores high FPS for those lovely bar charts and graphs. Even if it's still inferior to a single chip card (IMO) while being faster.
Regards,
SB
Consider the scenarios where your "flagship" (First to market) card is a large monolithic targetted at the relatively small enthusiast market and the case where your card is relatively small and targetted at the larger mainstream/performance mainstream market.
In both cases you'll have a relatively long developement cycle, that's unavoidable.
In the first case however you have a very large chip that's expensive to make, more prone to defects in manufacturing, and more prone to errors due to the sheer size and complexity.
In the second case, all those factors are mitigated to an extent. Less expensive, less prone to defects per die, less transistors to have a possibility of screwing up.
If you fail in the first case it's an extremely expensive failure with a low possibility of recouping R&D costs. Look at R600 as a prime example.
If you fail in the second case, it's less costly and you have a bit more wriggle room to recoup lost R&D through a larger market and an already low price point.
In the first case, success at best brings you a halo effect for the rest of your lineup and a boost to company image. Profits/revenue still won't be as good as the larger mainstream lineup. However, if the halo effect/image can boost your lineup enough that's moot. Problem is, in many cases the mainstream parts have tended to lag the halo parts by quite a few months the past few years.
In the second case, success hits the sweet spot of the market where the most money is made. It can also arguably cause a halo effect of it's own as you have the possibility of reaping numerous price/performance awards. Company image is further boosted in this important market segment as you earn word of mouth of providing best bang for the buck. And market can always spin that the company with the faster halo enthusiast part doesn't care about the masses and only caters to the elite rich boys, yada yada yada.
In other words. There's a lot of downside to targetting the enthusiast market as your main target and then moving down from there. And not a lot of immediate upside. Even if you execute well consistently.
Focusing on the mainstream/performance mainstream doesn have a whole lot of downside and while the upside isn't nearly as flashy it can be more profitable. As long as you don't forget about the PERFORMANCE.
Making a mainstream part and then not having leading or competitive performance will kill you. Just look at S3 which is still trying to get that right.Unless something drastic happens. I don't see ATI moving away from trying to hit the affordable performance segment in favor of the all or nothing enthusiast segment anytime soon.
And as long as sites benchmark as they have in the past. I don't see them moving away from multi-chip cards for the enthusiast market (even with icky AFR) as it still scores high FPS for those lovely bar charts and graphs.
Even if it's still inferior to a single chip card (IMO) while being faster.Regards,
SB
They did make specific decisions to target the performance segment but they had no idea GT200 would be so lackluster in comparison. In fact according to the article, they expected it to be more formidable than it turned out to be.
SB makes several good points about the risk/reward tradeoff of targeting the performance segment. But we have only anecdotal evidence (one generation) of how that actually works out in practice. It remains to be seen if that's going to be a trend or if AMD will be forced to go big again in the future in order to keep up.
For me the most important take-away from all this is that if Nvidia expects AMD to continue along this path it forces their hand in bringing new architectures to market in the $200-$300 segment much faster than usual and that's better for all of us.
Similar threads
