AMD Cayman info (or rumor)

[Oztopher]

There's been many rumors of the performance of this card, but it all points to being at least on par with a 580GTX. (anything less wouldn't make sense, for AMD, and judging from specs.)

Lets all just wait, honestly. All this talk of 30-40% quicker may very well lead to a disappointing and mediocre looking card.

[zalbard]

Lets all just wait, honestly. All this talk of 30-40% quicker may very well lead to a disappointing and mediocre looking card.

I agree.

Well that guy have 7k posts, so i don't think he is too stupid to give BS info.

Do you really believe in that? Ask Saaya if he ever posted any BS by mistake.

[bill_d]

hmm 30% to 40% and for maybe 399 to 449, can't be true it would be way too much fun
nvidia would go ape s**t

[flopper]

I believe.
The truth is out there.

[paulhamm]

Just a bit of fun with numbers. Caymen uses 90% of the die area for 4sps that Cypress uses for 5sps. Check the slides if you doubt. I used barts as a reference.
Cypress 334mm2 1600sp 256bit mem
Barts 255mm2 1120sp 256bit mem
Caymen ???? 1920sp 256bit mem
GF100/GF110 550mm2 512cuda 384bit mem
GF104/(GF114) 368mm2 384cuda 256bit mem

Die area estimates just for fun. Yep they have long WAGgy tails.
Caymen 334mm2/(1600/5)*(1920/4)*.9=450.9mm2
Caymen 334mm2/1600*1920=400.8mm2
Barts 334mm2/1600*1120=233.8mm2
Better Barts 334mm2/1600*1280=267.2mm2 What if HD6870 is a cut down Barts?
Barts reference accuracy 233.8mm2/255mm2=0.9168
Better Barts reference accuracy 267.2mm2/255mm2=1.0478

Barts is a significant refinement over the previous evergreen products. In the HD6870, Barts generates about 93% of the performance of the HD5870 Cypress while using less than 73% of the shaders. The question is how much of the gain in efficiency came from modified sps and how much came from a redesign of the uncore? It does seem clear that Cypress was sp heavy. That is Cypress has more sps than it can effectively use. Barts is the result of examining the evergreen products and finding ways to more effectively use the whole of the gpu.
Caymen XT should have an effective 50% performance advantage over a similarly clocked Cypress excluding performance enhancements to the uncore. AMD is showing in the various slides that the uncore changes in Caymen are even greater than those in Barts. Another question,,, is 4d=5d? We do not know, though AMD likely does. It does seem that 4d must be at least 90% as fast as the old 5d and likely more. The reason is that AMD only gains 10% die area with the change. If they did not see a potential gain then making the change would be valueless. I believe 40% better performance than the HD5870 would be a rather conservative estimate for the Caymen XT (HD6970). Depending on how well AMD has addressed the bottle necks of the evergreen (Cypress) gpu it could be that the HD6970 will be 60-70% faster than the HD5870. Pricing will be commensurate with actual performance. Expecting anything else would be silly, AMD is not a charity.
Should not be more than 3 weeks and we will know.

after thinking about it again i really expect 6970 to beat gtx 580 (those 30% might be possible in selected tests)


why? going from 5870 to 6970 we have the following improvements:

320 -> 480 shader clusters (50% improvement)
higher clockspeed
doubled front end with higher efficiency
2-3 times higher tesselation performance
doubled ROP performance
extra TMUs (5870 wasn't tmu bottlenecked)
2gb vs 1gb ram
6 (possibly 7) gbps ram vs. 4.8 gbps, possibly a reworked RAM controller with more cache = 25-50% more mem BW

everything in this list points to a performance improvement higher than 30% which puts 6970 on the same level as gtx 580 and that's a low estimate
additionally an improvement under 30% (= less performance than GTX580) would put 6950 at the same performance level as 6870 which makes no sense at all...)

(please tell me if i made a mistake in this list)

I can play crude math game too : if we take 6870 as a baseline,we have 1920 vs 1120 SPs.If we assume that the new 4way arrangement brings 90% of 5way performance in games and if we figure in the clock difference(that is in 6870 favor,900Mhz vs likely 800-850Mhz for 6970),plus a better mem BW(10% perf. increase impact) for 6970 we get : 1920/1120*0.9*850/900*1.1=1.6x faster than 6870. According to the image above, 70x1.6=112pts or 12% above 580. That's reasonable,12% I can buy. But 30-40% above 580 is highly unlikely IMO.

So the mighty Oracles have spaketh. Bow down in homage or face your DOOM!

paulhamm HD6970 40%-70% faster than HD5870 = faster than GTX580
generics_user HD6970 faster than the GTX580
informal HD6970 12% faster than GTX580

Stay tuned for our next episode when we will find out if the Oracles spaketh truth! Will they receive the adulation of the horde or be pilloried. How many nits will they and the rest of the horde of feral carnival monkeys pick? How much fecal matter will be thrown? For answers to all of these questions and more. You must watch the next exciting episode of "As the Oracles Spaketh" (Brought to you by our friends at "Soapy Suds" for all of your sudsy needs.)

[papatsonis]

i dont think the 6970 will be much faster (if ever...) than the 580.. the "crude algebra" which was pointed before, gives us 1920/4=480SPs*900mhz / 512SPs*772 = 1.0929 ~+9.3% theoretical advantage, if add the rumor about having trouble for target frequencies, and settle for 850mhz.. then its 480*850 / 512*772 = 1.0322 or a mere 3.2% theoretical advantage.. which in actual performance bonus will be negligible. (In fact if comes out with 850mhz clock i strongly believe it may be a tiny slower though..)

[generics_user]

i dont think the 6970 will be much faster (if ever...) than the 580.. the "crude algebra" which was pointed before, gives us 1920/4=480SPs*900mhz / 512SPs*772 = 1.0929 ~+9.3% theoretical advantage, if add the rumor about having trouble for target frequencies, and settle for 850mhz.. then its 480*850 / 512*772 = 1.0322 or a mere 3.2% theoretical advantage.. which in actual performance bonus will be negligible. (In fact if comes out with 850mhz clock i strongly believe it may be a tiny slower though..)

only problem with your math is that GTX580 shaders run at over 1.5ghz

[papatsonis]

indeed.. but with one more sub-"problem" to this logic.. a 480 compared to 5870 with your logic , should be 480*1404 / 320*850 -> 147% faster.. which ISNT the case.. while its much much closer to 480*702 / 320*850 -> 23.9% faster.. . I have no idea which part in Fermi acrhitecture is clocked to what speeds.. but the "least fast" part is what defines the final speed. Maybe it can brute force compute at 1404 mhz (in 480 example) but in graphics scenario.. has to wait for the 702 clock.. and the number in this post.. incline to this..

[Kaboka]

Didn’t some old rumor state that the 69xx might have a small size of memory directly on the core just like the xbox360 GPU? How will this factor into the performance or was this rumor killed?

[Heinz68]

When it comes to performance estimates I prefer to be conservative and if the product turns out to be much faster that's a nice surprise.

I think if the HD 6970 with smaller die size can be as fast as GTX 580, that is a win for AMD and if the HD 6970 turns out to be faster than GTX 580 that's a big bonus for AMD

[informal]

Hahahaha paulhamm,that was one funny post

PS I hope we don't get fecal matter on our forum avatars in the next episode

[ichime]

after thinking about it again i really expect 6970 to beat gtx 580 (those 30% might be possible in selected tests)


why? going from 5870 to 6970 we have the following improvements:

320 -> 480 shader clusters (50% improvement)
higher clockspeed
doubled front end with higher efficiency
2-3 times higher tesselation performance
doubled ROP performance
extra TMUs (5870 wasn't tmu bottlenecked)
2gb vs 1gb ram
6 (possibly 7) gbps ram vs. 4.8 gbps, possibly a reworked RAM controller with more cache = 25-50% more mem BW

everything in this list points to a performance improvement higher than 30% which puts 6970 on the same level as gtx 580 and that's a low estimate
additionally an improvement under 30% (= less performance than GTX580) would put 6950 at the same performance level as 6870 which makes no sense at all...)

(please tell me if i made a mistake in this list)

Wouldn't it be better to compare the shader clusters to the 6870 because otherwise, you're comparing the relatively inefficient Cypress clusters to the allegedly more efficient shader clusters in Cayman.

So lets say AMD maintains their shader efficiency in Cayman in their 4D cluster design and that 1 4D cluster does the same work as 1 5D cluster. Using the 6870's 1120 SP we get 224 shader clusters (1120/5) compared to the 6970's (1920/4) = 480 clusters. So from that standpoint, the 6970 has more than double the 6870's shader cluster count and should have twice the shader power (if the whole 4d = 5d thing holds true).

This is not to say that the 6970 will be twice as fast as the 6870 because despite the other enhancements, memory bandwidth is still limited and not all parts of the Cayman architecture are 100%+ increases over Barts.

I do believe it'll be well over a 60% increase over the 6870 in overall performance. If not then AMD wouldn't be successful in charging even $450 for this card.

[cold2010]

Catalyst 10.11 beta 8.790.6

Supports Cayman

"AMD Radeon HD 6900 Series" = ati2mtag_NICayman, PCI\VEN_1002&DEV_6718
"AMD Radeon HD 6900 Series " = ati2mtag_NICayman, PCI\VEN_1002&DEV_6719

[SicKlown42012]

Didn’t some old rumor state that the 69xx might have a small size of memory directly on the core just like the xbox360 GPU? How will this factor into the performance or was this rumor killed?

It's been said that Cayman has an off-chip buffer for tessellation. We don't know how exactly that will effect performance, but AMD wouldn't waste space(and drive up cost) if it didn't benefit it noticeably.

[MirageSys]

It's been said that Cayman has an off-chip buffer for tessellation. We don't know how exactly that will effect performance, but AMD wouldn't waste space(and drive up cost) if it didn't benefit it noticeably.

However the inclusion of the sideport to the R700 shows that they are willing to leave a design intact even without substantial (or noticeable) benefit. I wouldn't count on the tessellation buffer giving that much of a gain, but its certainly possible.


I would think it would have an impact on minimum framerates above all else, but I could easily stand corrected.

[Loque]

http://twitter.com/CatalystMaker

"Catalyst 10.12 is going to be HUGE"





will be huuuuge, like 10gbs :p

[tbone8ty]

wow so cat 20.10z they will get it optimized?

[Chumbucket843]

i dont think the 6970 will be much faster (if ever...) than the 580.. the "crude algebra" which was pointed before, gives us 1920/4=480SPs*900mhz / 512SPs*772 = 1.0929 ~+9.3% theoretical advantage, if add the rumor about having trouble for target frequencies, and settle for 850mhz.. then its 480*850 / 512*772 = 1.0322 or a mere 3.2% theoretical advantage.. which in actual performance bonus will be negligible. (In fact if comes out with 850mhz clock i strongly believe it may be a tiny slower though..)

your calculations are way off, even for a crude estimate.

the theoretical advantage for ATi's shader core is still about 2x.

[Sn0wm@n]

the wait is getting impossible ......

http://twitter.com/CatalystMaker

"Catalyst 10.12 is going to be HUGE"





will be huuuuge, like 10gbs :p

didnt you read about the same size of the current package ???? so i doubt that it will be that big ....

[zerazax]

One more week...

[xBanzai89]

One more week...

I'm hoping the 8th like some rumors aid .

[Tuvok-LuR-]

based on my calculations 6970 should be THIS faster than 580:

[zalbard]

Stock support for GPU + APU, looks like Bulldozer will be a way to go for the highest graphics performance!

[96redformula]

Hm...So GPU power is additional with a GPU+Fusion CPU.

[Tomasis]

sweeeet