AMD: R9xx Speculation

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.

 

know whats funny. Those of us in the know , know that there is no Cayman or 6970/50 . No this is all fud from AMD so that nvidia thinks its still on 40nm . We all know that they decided to launch the 7970 on 32nm and it offers 5870 in quad fire performance with only a 100w power usage and the 7930 verison will come with a passive cooler with prices starting at $200-$400.

 

Did you miss out on the 6850 boards that had 1120 shaders available that some reviewers got?

http://www.bjorn3d.com/read.php?cID=1949
http://www.tweaktown.com/articles/3...eo_card_w_1120_shaders_overclocked/index.html
http://www.geeks3d.com/20101025/rad...aders-tessmark-and-shadertoymark-comparisons/

So yes, AMD have dropped a big 'misinformation' bomb shell.

 

To be fair, you are projecting your way of thinking to AMD, I doubt you understand their mindset. From many interviews I've read and basing on their strategy, It seems AMD's goal is a sweetspot, and not to be Top at any cost, i.e. medium sized chip with a very good price/performance, which would be profitable. They reached those goals since the 38xx days.

For the Top card, AMD also was Top for over a year now, and before - briefly with 3870 X2 and 4870 X2, since they adopted new strategy. Speaking of single GPU only, AMD could have easily remained in the Top with 5870 even after 480 launch (binned and OC'ed 5890 range would be just fine), but AMD refused to do that. Why? Because its not their goal to be the fastest with single-gpu at any cost, hence its yet another proof of what I'm saying.

 

Then, there are some issues...

I think it's your own projection.

1- Barts is very similar to RV770 in terms of size, cost and pricing (and if you look closely, it's horribly worse in the "normalized perf/mm²" department, it should be ~200mm²)

2- Barts ended up in the "6800" series, Cayman is "6900". Cayman therefore has to be able to pretend to a $400+ price tag.


Now, if one year later than Cypress "not sweet-spot but still not enthusiast" based HD5870 you brand something as low as 25% faster at most as "enthusiast" (best case if HD6970 < GTX580 unless using a biased benchmark panel), something went very wrong.

 

Seems to be an oddity/mistake not a disinformation campaign. Dont see the big sites getting these either.

 

PSU-failure, show me AMD claim they want to be Top at any cost with a single GPU, you wont find anything for this, yet for the sweet spot you'll get 1 mln.+ hits, hence my post and it has nothing to do with "my projection".

Neither of your points proves the idea of AMD's goal to have Top single GPU at any cost, and if you want to attack their sweet-spot strategy and GPU gens, pester Dave about it From my point of view its working fine, actually its brilliant to the extent I'm surprised Nvidia havent adopted it yet (just check both companies strategies execution the last couple of years), even when NV itself said "its f*** hard to build such huge chip".

 

Still a few days to go but I think , Nvidia's will win this year on performance for a single chip and AMD will win on price/performance as per the past year.

AMD will by the most expensive and fastest card(5970) which I tend to hate because I'll never buy a card like this.

Nvidia wins the team of the year for me for delivering the 580 and 570 before AMD could deliver their cards.

 

GTX570 was launched 1 week before Cayman. That's hardly a win. Maybe somebody could appreciate, that GTX580 was launched 1 month before Cayman... but I can't imagine why should one month be so important in current situation. Half of the high-end users waited more than six months for GTX480... and those, who decided to buy it, will hardly spend $500-600 for 15% performance and those, who decided to wait for additional six months probably won't care for 3 or 4 weeks.

 

I commend Nvidia for getting the 580 and 570 out, but keep in perspective it's a refresh (or, rather, a card that should have been out a year ago at this point) and/or fix of an existing card, whereas AMD pushed out another generation more or less on schedule (once every 14 months or so)

 

1) Yet run dx 11 features on a game and tell me how well the rv 770 does against barts ?

2) Why ? What happened to the $250-$400 price points ? Doesn't amd want products there ? When nvidia was still trying to bs having a dx 11 card in 2009 amd already was launching a top to bottom solution . Last i recalled it toped out at $380 bucks with a dual gpu card covering the very high end.

Why can't it be that way again ? Barts in the $150-$250 price range and Cayman in the $300-$400 range and Cayman dual gpu at the $600 range.

If its the 580 your worried about cf 6870s are already faster , so even dual 6950s will be faster in a single card.

As for worrying about what should be what , if oyu want to bust on amd for naming , look at nvidia's line up. Is the 580 a generation above the 480 ? Both companys equaly suck here

 

Click!

The note in the picture says: "It's real!".
The comment in the post: "Mostly dissapointed for now".

Update:

More -- transaltion: "In the higher resolutions (>1680*1050) it surprasses 6870 by 30-40% in the best case, but it doesn't really justify an upgrade...".

More info to come, soon!

 

Looks like they've got a 6950. Can you translate their follow up post:

My translation says at higher resolutions, it can reach 30-40% faster than the 6870 in best scenarios. Interestingly enough, 30-40% faster than 6870 in best case scenarios (games that AMD is doing well in) puts it right under the 580 and above 570, which isn't exactly bad for a 6950...

Also, did they say what they were testing?

Because I've noticed a lot of people complain about performance, but often it comes from the 3dMark benching crowd

Then I thought about it, and it makes sense:

1408 * 800 * 2 = 2.252 TFlops
1536 * 880 * 2 = 2.703 TFlops

Compared to Cypress:

1440 * 725 * 2 = 2.088 TFlops
1600 * 850 * 2 = 2.720 TFlops

In terms of pure math, Cayman is more or less a wash against Cypress. With synthetics, driver optimizations and scoring algorithms certainly wouldn't help Cayman which in the eyes of the test, is more or less the same shader output - TMUs and ROPS and improvements to tesselation will help, but how much? I know pure math isn't everything of course, but it is certainly one of the key tests in the benchmark suites

The Chinese boards were awash with (badly translated) statements about how it's disappointing in 3dMark, but real world performance puts it closer to the 580 than the benches indicate

edit: I see you did translate it, so the translation was correct

30-40% faster than the 6870 in best cases puts it ~20% faster than the 5870, which is definitely lower than people think, but that puts it ~480/570 territory and so the 6950 should be ~580 territory

Not 5970 performance, but then again, a < 400mm^2 chip getting 580 levels of performance makes me wonder what Antilles can do

 

As I said, because sweet-spot does work for a mainstream/performance part (regular gasoline in a regular car) doesn't mean it'll work for a "competition" part (same regular gasoline in a race car). The marketing name itself says Cayman is an Enthusiast product, and as such it doesn't obey the same rules.

Either you design a product for low cost, or you design it for high performance, Cypress was in-between and a little too biased towards low cost already, which led to Barts not being that slower with 30% less SIMDs and ~25% less die size, that's an obvious example of counterproductive savings.

 

Looks like the poster ninja-edited the % number's on the BG website

While Cypress was indeed an in-between, you can't really say that Barts being more efficient is because of Cypress being an in-between. There could have been serious design flaws in Cypress that were only discovered later with Barts + process improvements and chip re-balancing. Basically what the GTX 580 / 570 did

Remember that Japanese analysis of Cayman that showed a clear progression in die size & performance for AMD's high end single GPU

Taken from another perspective:

RV670 - 192mm^2 - 3870 was 65% of 9800GTX
RV770 - 255mm^2 - 4870 was 75% of GTX 285
Cypress - 334mm^2 - 5870 was 85% of GTX 480
Cayman - ~385mm^2 - 6970 will be xx% of GTX 580

You can argue the %'s if you want, but it's the general trend, esp. if the latest 570 reviews are correct and the 5870's driver improvements place it closer to 90% than 85%

Interestingly enough, each progression up to Cayman was ~30% larger than its predecessor - Cayman breaks that to being only 15% larger if the 385mm^2 rumor is true

However, performance figures suggest Cayman as being the closest one yet to the competition

If the 6xxx generation was where the high end and the sweet spot split, then it's a good start - the next generation, AMD may well push for the full high end win given the trend

 

Now, this is suppose to be a "really" real slide, just to iron the ALU count rumors.

 

IT's the same. this is the last slide of the so called "fake deck"

 

So 1536 is the fake number ?

 

In light of all the new rumours, can we revisit the one-month delay again. I still can't comprehend why AMD delayed the cards. Is it because?

1- Drivers were broken

2- Card wasn't performing well enough in 3D mark 11

3- AMD knew Cayman XT won't beat GTX 580 and was waiting for GTX 570 to release so that it can price 6970 and 6950 depending on Nvidia prices.

4- (Fanboy dream) GTX580 release did not impress AMD and it purposefully move to cripple Cayman. The company knew that 40nm is going to last for another year and it may require a full Cayman with 1920sp by middle of the next year to encounter anything that Nvidia may bring.

 

Is it possible we've just hoped for too much, and AMD have done the same as they did last time around?

This would be a double chip at the top end to take the fastest single card (5890/Antilles), single chip a little below Nvidia's top single chip (5870/6970), but much smaller, cheaper, less power and heat in the sweet-spot that makes AMD a lot of profit. Underneath is a massively good bang for the bug lower performance range for the salvage parts (5850/6950)

Maybe we're going to see exactly the same strategy of AMD forgoing the single massive monolithic big chip that's difficult and expensive to make in favour of smaller, efficient chips that have high profits, that can be priced very competitively against Nvidia and provide almost as good performance.