AMD: R9xx Speculation

[Grall]

If you're in the camp of Cayman XT competing with the $350 570 GTX being AMD's real performance target and believe that's not a disappointment, then you either weren't posting to this thread or you had a sudden, blinding, revelatory experienced that wiped some memory banks.

Don't project your views and opinions onto others, kkthx.

I don't see why anyone would seriously expect AMD to shift focus overnight (or, well, in the space of a year, really) from building stuff that's cheap and good enough to world-beating cutting edge performance, and getting there with a chip that's alledgedly only 3/4 of their main competitor's offering.

Nothing's free in 3D. If you're only 3/4 the size, you're unlikely to beat your adversary, and certainly not by 50%. Not unless the competitor (Nvidia, in this case) is incompetent to a degree not seen in this business for many years now.

 

[Grall]

But there is a reason why the X2 card is now the 6990: Sideport is back. There is that off-die buffering and scaling is sustained at near linear without the need for profiles. Also, RAM is shared, so frame buffer is not duplicated anymore. Also, this smokes the 580 silly...

Yes, and I can also reveal it turns water into wine, cures world hunger and brings peace on earth.

Those features rely on future driver updates however that may or may not materialize due to hardware bugs found late in development, and were in fact the genuine reason for this recent delay...

 

[Mize]

Don't project your views and opinions onto others, kkthx.

I admit it! There's some self projection there as I was hoping for a beast.

 

[Grall]

Hey, I was hoping for a beast too, but I'm OK with cayman being 20, 30% or whatever slower than the 580 as long as it's 20, 30% cheaper, cooler and quieter. Add new power management stuff, better GPGPU capabilities, fixed aniso filtering finally (they BETTER get it right this time god dammit!), new AA modes and so on, it could still be a very appealing offer.

Especially as the 580 has VERY low availability. In Sweden there's basically only 3 (major internet) retailers that have it in stock right now according to the pricewatch website I'm using, and it's not particulary cheap either. And only a handful different vendors' models available also I might add.

 

[fbomber]

Anand has Metro 2033 570=10% faster than 5870. Civ 5 <10% Ditto Crysis Warhead. I didn't study the issue but just by glancing at Anand's review I got the feeling 5870 was a lot closer to 570 in more modern games than older ones.

http://www.techreport.com/articles.x/20088/9
Metro 2033

Very high
5870: 22
GTX570: 27
Difference: 22%

High:
5870: 31
GTX570: 40
Difference: 29%

Medium:
5870: 35
GTX570: 54
Difference: 54%

Now look at the max settings of the other games I mentioned.

http://www.techreport.com/articles.x/20088/5
Lost Planet 2
5870: 21
GTX570: 31
Difference: 47%

http://www.techreport.com/articles.x/20088/11
Dirt2 DX11
5870: 80
GTX570: 100
Diiference: 25%


The difference from 5870 to GTX580 is a lot higher. All I´m trying to say is that 6970 must be much faster than 5870 to be faster than GTX580 across the board, specially at the DX11 games.

 

[chavvdarrr]

I think that ATI have never intended to target Cayman against GTX580. That should be Antilles's job (they wouln't have Antilles in roadmap if they planned to beat GTX580 by single-GPU product). But GTX580 is only bug-fixed/15-20% faster Fermi, so GTX580 performs closer to Cayman than to Antilles. That's in fact a better result than they expected and the postponed launch of Antilles just proves it.

Gf104b^2, cough.

 

[no-X]

I think chavdarr is correct in saying that Cayman puts ATi in a relatively worse position.

I don't think so. Cypress wasn't able to deliver great perf/price around $200-250 and there was still quite significant gap between HD5870 and GTX480 which was even worse with tessellation enabled. HD5830 was too slow in its segment, HD5850 was fine and HD5870 was quite feature/performance-lacking in its segment. Now ATI targeted $200-250 segment by Barts, which offers more performance (with 100mm² smaller core) than GF104. High-end segment is targeted by a GPU with better AF, better tessellation performance and some nice (near luxury) features - simply features, which are important for high-end product. I'd say, that they are in better position: They covered more market's segments with products, which meet customers' requirements in much better way.

Nvidia have managed to improve performance of Fermi without increasing die size

In fact they didn't improve performance of the GPU - they just enabled previously disabled part. It would be like saying, that HD6970 is HD6950 which's performance was improved without increasing die size

If ATi can increase the clocks to compete with GTX580 directly with 6970 then it puts them in a strong position, a smaller die will be cheaper and very popular.

Did they need to overclock RV770 or Cypress to competitiors performance levels to get in good position?

I think the most iportant thing is price. The only problem I see is that price/performance can be negatively affected by 2GB's of GDDR5...

 

[NathansFortune]

I don't think so. Cypress wasn't able to deliver great perf/price around $200-250 and there was still quite significant gap between HD5870 and GTX480 which was even worse with tessellation enabled. HD5830 was too slow in its segment, HD5850 was fine and HD5870 was quite feature/performance-lacking in its segment. Now ATI targeted $200-250 segment by Barts, which offers more performance (with 100mm² smaller core) than GF104. High-end segment is targeted by a GPU with better AF, better tessellation performance and some nice (near luxury) features - simply features, which are important for high-end product. I'd say, that they are in better position: They covered more market's segments with products, which meet customers' requirements in much better way.

Yes, that's all true, but in the past now. The discussion is about 6900 series...

In fact they didn't improve performance of the GPU - they just enabled previously disabled part. It would be like saying, that HD6970 is HD6950 which's performance was improved without increasing die size

Semantics. For the end user perfomance has increased by around 10-20% for the same die size and better thermals. It's a new product.

Did they need to overclock RV770 or Cypress to competitiors performance levels to get in good position?

Yes? Is that a trick question? ATi parts been clocked significantly higher than their Nvidia counterparts to get competitive performance from the smaller die.

I think the most iportant thing is price. The only problem I see is that price/performance can be negatively affected by 2GB's of GDDR5...

Indeed. 2GB of 5.5GHz GDDR5 won't come cheap, and the increased die size vs GF110 will hamper profitability too whether you see it or not.

 

[Bouncing Zabaglione Bros.]

Nvidia have managed to improve performance of Fermi without increasing die size or power by around 10-20% while ATi have only managed a similar improvement by increasing the die size by a similar amount.

Not surprising given all the problems and the cutdown 480 that was released last year. There was lots of room for Nvidia to improve their 40nm products. Much less so with ATI who got it pretty right straight away.

 

[leoneazzurro]

Yes? Is that a trick question? ATi parts been clocked significantly higher than their Nvidia counterparts to get competitive performance from the smaller die.

Are you forgetting that the shader clock of Nvidia cards is way higher than AMD cards since G80?

 

[NathansFortune]

Not surprising given all the problems and the cutdown 480 that was released last year. There was lots of room for Nvidia to improve their 40nm products. Much less so with ATI who got it pretty right straight away.

Of course. Regardless of whether GF110 or GF100b is or isn't a completely new product the end user has a new GPU in the GTX580 or 485 as it should have been called and it delivers 10-20% better performance for no increase in die size. It means that Nvidia will keep their single GPU crown. For the end user, GTX480 > 5870 and GTX580 > 6970 but to keep up ATi have increased the die size of the 6970 by 20% while Nvidia haven't for the same performance disadvantage as before.

To sum it up it has gone down like this:

GTX480 @ 520mm^2 15%> HD5870 @334mm^2
GTX580 @ 520mm^2 15%> HD6970 @390-400mm^2

So objectively ATi are in a worse off position than before. Whether that is down to GF100 being so bad or Cypress being so good it makes no difference. History is exactly that. Obviously this is speculation based on rumours which may or may not be reliable, but as of now it looks like this is the most likely scenario.

 

[CRoland]

You know, I find it simply strange that, now that we're reasonably sure Cayman can't go toe to toe with the 580, people are arguing that AMD never had that as a goal and that nobody would ever expect that from a smaller chip.

The speculation hasn't really been based on AMD's targets, only what people thought or hoped they would be. When Cayman launches we'll see if:

1. Cayman launches with fully enabled units and high clocks (Cypress level or higher) and still doesn't perform as well as 580. That implies AMD probably didn't target much higher performance for whatever reason.
2. Cayman launches with disabled units or low clocks and loses to 580. AMD may have tried to produce a monster GPU, but failed.
3. Cayman actually trounces 580 and AMD guys are very happy.

 

[no-X]

Yes, that's all true, but in the past now. The discussion is about 6900 series...

That's not past. They don't need target mainstream segment with Cayman, because it's Bart's job. So Cayman can be slighlty more high-end oriented and target its segment significantly better than Cypress.

Semantics. For the end user perfomance has increased by around 10-20% for the same die size and better thermals. It's a new product.

End user doesn't care about die-size. If you care about technological details like die-size, you can't ignore, that GTX480 was a partially disabled part. You are comparing die-size, but you forgot, that die-size is important primarily because of yields. Does fully enabled and higher-clocked GTX580 yield better than partially disabled/underclocked GTX480 to translate the hypotetical advantage into a real one? Some forum members don't think so...

Regarding Cayman we know, that performance was boosted more than die-size, so manufacturing costs / performance ratio is very likely better.

Yes? Is that a trick question? ATi parts been clocked significantly higher than their Nvidia counterparts to get competitive performance from the smaller die.

Hot-clock got ignored?

Anyway, back to topic. Cayman has about 5% bigger die than GTX460 and performs about (guess) 5-10% under GTX580. I cannot imagine what better could ATI wish? Theirs GPU has manufacturing cost comparable to competitor's mainstream and performs almost like competitor's high-end. Without the need of more advanced manufacturing process, without lack of any important features, without significantly higher clocks and with a bunch of new features.

 

[NathansFortune]

If we follow that through (and I thought GF104 was actually measured at ~335mm^2 not Charlie's 370mm^2) then the GF114 with all SMs enabled for a 384SP chip will give a 20% boost taking it to 6950 levels for a smaller die size.

That kind of comparison is flawed because we don't know the cost structure of each company, I mean Nvidia made money even though all the evidence was pointing at GF100 being a massive loss-leader while ATi lost money even though they had the whole DX11 market to themselves with Cypress.

 

[PSU-failure]

Anyway, back to topic. Cayman has about 5% bigger die than GTX460 and performs about (guess) 5-10% under GTX580. I cannot imagine what better could ATI wish? Theirs GPU has manufacturing cost comparable to competitor's mainstream and performs almost like competitor's high-end. Without the need of more advanced manufacturing process, without lack of any important features, without significantly higher clocks and with a bunch of new features.

I can imagine, easily.

In that case, going for a slightly bigger die wouldn't have such an impact on yield nor on die candidates per wafer, but that would have given them the performance crown.

Performance crown itself allows for a price premium of at least $50 per board, and brand recognition would have increased too.


Again, that's probably the biggest difference between a "pure economics" strategy (what some call "sweet spot", although it isn't... each segment of the market has its own sweet spot) and good product positioning.

 

[no-X]

If we follow that through (and I thought GF104 was actually measured at ~335mm^2 not Charlie's 370mm^2)

Really?

then the GF114 with all SMs enabled for a 384SP chip will give a 20% boost taking it to 6950 levels for a smaller die size.

Really?

That kind of comparison is flawed because we don't know the cost structure of each company, I mean Nvidia made money even though all the evidence was pointing at GF100 being a massive loss-leader while ATi lost money even though they had the whole DX11 market to themselves with Cypress.

Really?

ATI Q4 2009 $58M
ATI Q1 2010 $47M
ATI Q2 2010 $33M
ATI Q3 2010 $1M

When exactly they lost money? Only nVidia lost $141M in Q3/10 (Q2/11 FY)

 

[jimbo75]

They've both made $81m this financial year actually.

 

[NathansFortune]

Really?

I must have remembered it wrong.

Really?


Really?

Err, I don't see any performance indications for a 384 SP GF104 in that chart, perhaps you can point it out for me...

ATI Q4 2009 $58M
ATI Q1 2010 $47M
ATI Q2 2010 $33M
ATI Q3 2010 $1M

When exactly they lost money? Only nVidia lost $141M in Q3/10 (Q2/11 FY)

Fair enough!

 

[leoneazzurro]

Err, I don't see any performance indications for a 384 SP GF104 in that chart, perhaps you can point it out for me...

A GF104 with 384 SP will have 15% more shader power. How this couls translate in a 20% more performance in real games without a massive clock increase (and this means lower yields) is already debatable, but if 6950 is 15-20% more powerful than a 5870 then this "full GF104" will be not enough to reach it.

 

[no-X]

Exactly. 20% over GTX460 wouldn't be enough to beat HD6870... I don't expect Cayman to be slower than Barts :smile: