AMD Cayman, Antilles Specifications Surface

[bear jesus]

The memory bottleneck was a myth on the 5xxx series. You could over clock the memory on my old 5970 from 1000 - 1200 and gain a few percent performance. Overclocking the GPU on the other hand form 725 - 900 was a HUGE jump in performance. The problem with the cards was a lack of memory on very res with AA. At 6050 x 1080 AA with a lot of games was not possible.

My 480's regularly use more than 1 GB in games at that res. Hell there is even a few games that use more than 1 GB at 1920 x 1080 with 4/8 x AA.

Overclocking the memory on the 480's does very little as well. Again OC the GPU massive jump in performance in most games.

Most of the games i can't max on a single card it's often the AA that can't be maxed at 5040x1050 and the same applies to when I'm using 2 cards so i assume it is due to running out of vram.

I really hope the 6970 is about the speed of 6870 crossfire (although only due to imperfect scaling thus 1920 sp's being about the same performance as 2240 in crossfire) but excels anywhere that vram is a limit, If so i will be saying goodbye to my 6870's as i have never been a fan of dual card setups, i just got these because i could not wait any more

 

[Aditya]

3840/2 = 1920

6970 would only have 1920 stream processor.
it couldn't beat GTX580.

i hope 6970 would be double of 6870.
this will surely beat GTX580.

i will stick with my 5870 E6.

But 6 series are meant to be value budget cards,so you cant blame them for not competing with a very costly card!

 

[NdMk2o1o]

But 6 series are meant to be value budget cards,so you cant blame them for not competing with a very costly card!

Sorry, where did you hear that?

 

[TheMailMan78]

Did anyone notice this one has 2 polygons per clock vs 1 from the last gen?

 

[wahdangun]

Did anyone notice this one has 2 polygons per clock vs 1 from the last gen?

yes, so i hope thats mean each shader core was 2x the performance of cypress

 

[bear jesus]

Did anyone notice this one has 2 polygons per clock vs 1 from the last gen?

Yes but didn't have a clue what relevance it had, are the 68xx cards 1 per clock like the last gen as well?

Is it kind of like doubling the ipc for a cpu or something? I'm kind of curious what effect on performance it will have.

 

[TheMailMan78]

yes, so i hope thats mean each shader core was 2x the performance of cypress

Thats what I'm thinking. People get hung up on clocks and neglect the basics. We will just have to wait and see some benched. That and did you notice this bitch is 4 gigs!

 

[Lionheart]

I will take 2 HD6990's in Xfire thankyou

 

[Benetanegia]

The other slide.

http://www.rumorpedia.net/wp-content/uploads/2010/11/radeonhd6970.gif

These slides are looking fake to me by the moment. Like I said 30 SIMD and 96 texture units are not posible, unless the architecture is radically different and they're doing something like: 5 SIMDs are packed along with 16 texture units in some kind of cluster or something strange like that (which is not good for harvesting). The previous modular approach by AMD was "clusters" of 1 SIMD + 4 TMUs and this way they could disable each failing cluster "individually" in order to harvest the chips and make the HD5850 or 5830. The 5 SIMD + 16 TMU clusters would be highly innefficient when it comes to harvesting the GPUs as any error would render 1/6th of the chip unusable, but then again maybe they did that and that's why they are having so bad yields and had to delay the release.

Summarizing, I fail to see how 96 TMU can be evenly split into 30 clusters, so I think the slides are fake until proven wrong.

@ 2 polygons per clock

It only means that it can calculate and rasterize 2 polygons per clock. Nothing to do with shader performance. It will arguably have an effect on tesselation performance, but in theory, judging by specs and theoretical throughoutput, it shouldn't have any effect tbh: even at 1 poly/clock the previous gen (including Barts) should be able to perform 725-900 million polygons/s (like 12-15 million polys per frame @60 fps), which is much much more than any scene has even with superextreme tesselation. To make an idea, a 1920x1200 picture has 2.3 million pixels, so a 12-15 milion poly scene would have an average of 5-8 polys per pixel, that is, using the wireframe mode on Heaven benchmark would render a completely white screen. Not some parts appearing almost white like Heaven Extreme setting does, no, completely white. And it would happen almost the same even with 2.3 million polys which is 1/8 of Cypress theorethical capabilities. What I mean is that if the specs for Cypress were real or anything close to realistic (i.e actual throughoutput being 1/8 of the speced one), there should be no significant improvement, because Cypress/Barts shouldn't be bottlenecked to begin with.

 

[left4lol]

The other slide.

http://www.rumorpedia.net/wp-content/uploads/2010/11/radeonhd6970.gif

Those slide are fake this is the real slide


some other slide

 

[Fourstaff]

I have a question, is the 69xx architecture an extended version of the 68xx architecture or is it another experiment from AMD? Based on the details of the slides, it looks like the 69xx and the 68xx TMU/cluster is different (thanks for prompting me to do the search, Benetanegia).

 

[Benetanegia]

Those slide are fake this is the real slide
http://www.abload.de/image.php?img=fot_019udly.jpg
http://www.abload.de/thumb/fot_019udly.jpg
some other slide

http://www.abload.de/image.php?img=fot_019udly.jpg
http://www.abload.de/image.php?img=fot_022jedn.jpg
http://www.abload.de/image.php?img=fot_024ofv9.jpg
http://www.abload.de/image.php?img=fot_0216i96.jpg
http://www.abload.de/image.php?img=fot_023xhnq.jpg
http://www.abload.de/image.php?img=fot_0204h3x.jpg

LOL. Thanks.

Interesting thing in those slides is that AMD is going to contain TDP too and by the looks of it far more extensively than Nvidia did with the GTX580 (i.e Nvidia only does it on Furmark, etc.). I wonder what people will say about that. I'm prepared for the hypocrisy.

I'm not against those measures btw, and I think that the implementation shown in those slides is better than Nvidia's, mostly because it always works and because it can be tunned by advanced users. Fact still remains that average users might end up with reduced performance in some cases or with pointless/counterproductive overclocks. Mostly what everone feared about the feature in GTX580 looks like it's 10x worse on Cayman.

Also the fact that SIMD units are not yet determined does suggest that the rumors about bad yields were correct. GTX480 AMD version reloaded here we come... (maybe not, but that's what TBD suggests at this point (Oct 2010))

 

[TheMailMan78]

LOL. Thanks.

Interesting thing in those slides is that AMD is going to contain TDP too and by the looks of it far more extensively than Nvidia did with the GTX580 (i.e Nvidia only does it on Furmark, etc.). I wonder what people will say about that. I'm prepared for the hypocrisy.

I'm not against those measures btw, and I think that the implementation shown in those slides is better than Nvidia's, mostly because it always works and because it can be tunned by advanced users. Fact still remains that average users might end up with reduced performance in some cases or with pointless/counterproductive overclocks. Mostly what everone feared about the feature in GTX580 looks like it's 10x worse on Cayman.

Also the fact that SIMD units are not yet determined does suggest that the rumors about bad yields were correct. GTX480 AMD version reloaded here we come... (maybe not, but that's what TBD suggests at this point (Oct 2010))

Are you referring to the throttling?

 

[Swamp Monster]

Those slide are fake this is the real slide

I doubt this slide is real too, because of TBD of some values in it. Come on, it is already November 22!

 

[Benetanegia]

Are you referring to the throttling?

This:
http://www.abload.de/image.php?img=fot_024ofv9.jpg

and this:
http://www.abload.de/image.php?img=fot_023xhnq.jpg

. Clamps GPU TDP to a pre-determined level.
. Integrated power control processor monitors power draw every clock cycle
. Dynamically adjusts clocks for various blocks to enforce TDP
. Provides direct control over GPU power draw (as opposed to indirec via clock/voltage tweaks)
. etc...

 

[TheMailMan78]

This:
http://www.abload.de/image.php?img=fot_024ofv9.jpg

and this:
http://www.abload.de/image.php?img=fot_023xhnq.jpg

. Clamps GPU TDP to a pre-determined level.
. Integrated power control processor monitors power draw every clock cycle
. Dynamically adjusts clocks for various blocks to enforce TDP
. Provides direct control over GPU power draw (as opposed to indirec via clock/voltage tweaks)
. etc...

Yeah. I don't like the fact I have less control. Of course that wasn't a big problem on the 580 anymore then anything else. My issue with ALL these cards is I want to decide what I can and cannot do via voltage and clocks. ATI or Nvidia. If I want to dump water on one and knock the volts up to 3v then so F$#KING be it.

However I know why they are doing this. Every jackass on the planet now is an "expert" overclocker. This is just warranty control IMO.

 

[H82LUZ73]

At last, specifications of AMD's elusive Radeon HD 6970 and Radeon HD 6990 graphics accelerators made it to the internet, with slides exposing details such as stream processor count. The Radeon HD 6970 is based on a new 40 nm GPU by AMD, codenamed "Cayman". The dual-GPU accelerator being designed using two Cayman GPUs is codenamed "Antilles", and carries the product name Radeon HD 6990.

Cayman packs 1920 stream processors, spread across 30 SIMD engines, indicating the 4D stream processor architecture, generating single-precision computational power of 3 TFLOPs. It packs 96 TMUs, 128 Z/Stencil ROPs, and 32 color ROPs. Its memory bandwidth of 160 GB/s indicates that it uses a 256-bit wide GDDR5 memory interface. The memory amount, however, seems to have been doubled to 2 GB on the Radeon HD 6970. Antilles uses two of these Cayman GPUs, combined computational power of 6 TFLOPs, a total of 3840 stream processors, total memory bandwidth of 307.2 GB/s, a total of 4 GB of memory, load and idle board power ratings at 300W and 30W, respectively.

[url]http://www.techpowerup.com/img/10-11-22/153a_thm.jpg[/URL]

Source: 3DCenter Forum

Ok look at this slides date guys and cut half of the memory gb/s and you have 6970 those other slides look like a presentation for internal use why they are dated October .

 

[entropy13]

Yeah. I don't like the fact I have less control. Of course that wasn't a big problem on the 580 anymore then anything else. My issue with ALL these cards is I want to decide what I can and cannot do via voltage and clocks. ATI or Nvidia. If I want to dump water on one and knock the volts up to 3v then so F$#KING be it.

However I know why they are doing this. Every jackass on the planet now is an "expert" overclocker. This is just warranty control IMO.

You can change settings related to it in the AMD Overdrive section.

"User controllable via AMD Overdrive Utility."

 

[HalfAHertz]

The TMUs were linked to the ROPs not to the SPs last time I checked, so 96 could very well be true.
But to be honest I share your concerns about the low yields. If they were truly having problems with pcb components and not the chips themselves, they wouldn't have left so many things unspecified. Unless of course they're trying to mislead Nvidia about the performance figures? But that would be pointless at this point imo...

I am pretty sure AMD will be torn by the performance community about the TDP limiter, unless they make an easy opt-out solution in the drivers.

 

[Benetanegia]

You can change settings related to it in the AMD Overdrive section.

"User controllable via AMD Overdrive Utility."

Yup. I was going to tell him that.

The problem is for average users and most probably with factory overclocked cards. How is that going to work? Very few people use Overdrive, at least that I know of. And none of the non-enthusiast friends would ever use it, that's why pre-overclocked cards exist, because most people don't OC themselves or they only do the basic upping the clock slide and see what happens. That's all they want to care about.

And my concern is that by default the TDP limit will most probably be imposed for default clocked cards so OCed cards will probably exceed that limit and to make things worse, they will exceed that limit and get underclocked when the performance would be most needed. I mean when is TDP going to be higher? When the GPU is most used, obviously, so I'm almost sure that unless the TDP limit is automatically adjusted depending on the card in use and the clocks, most people are going to get massive slowdowns, because performance (clocks) will be crippled when GPU power was most needed, which is always when you are getting lowest frames. (lowest frames based on GPU, not lowest frames imposed by the CPU)

The TMUs were linked to the ROPs not to the SPs last time I checked, so 96 could very well be true.

No offense, but you didn't check very well or you didn't check since pre DX10 era. Ever since R600 TMUs have been linked to SIMDs

 

[TheMailMan78]

You can change settings related to it in the AMD Overdrive section.

"User controllable via AMD Overdrive Utility."

Missed that. Thanks. I fail.

Yup. I was going to tell him that.

The problem is for average users and most probably with factory overclocked cards. How is that going to work? Very few people use Overdrive, at least that I know of. And none of the non-enthusiast friends would ever use it, that's why pre-overclocked cards exist, because most people don't OC themselves or they only do the basic upping the clock slide and see what happens. That's all they want to care about.

And my concern is that the TDP limit will most probably be imposed for default clocked cards so OCed cards will probably exceed that limit and to make things worse, they will exceed that limit and get underclocked when the performance would be most needed. I mean when is TDP going to be higher? When the GPU is most used, obviously, so I'm almost sure that unless the TDP limit is automatically adjusted depending on the card in use and the clocks, most people are going to get massive slowdowns, because performance (clocks) will be crippled when GPU power was most needed, which is always when you are getting lowest frames. (lowest frames based on GPU, not lowest frames imposed by the CPU)

I think its going to be temp regulated much like the 580. I doubt it will be clocked down if the temp is 30c

 

[HalfAHertz]

The stupid thing here is that if you choose to "opt-out" and disable the limiter, you'll most likely void the card's warranty...

 

[mechtech]

Although, threaded AMD promises, but I have to agree this is incredible weak show from AMD.

Considering AMD/ATI is at huge issues with driver, BIOSes and all software of software issues + the fact that they have delay issues and nothing(as in new games) never seems to work correctly on their GPUs it is pretty funny..

(and nobody dare to accuse me an 'NVIDIA' fanboy I own only AMD/ATI GPUs atm a lot of them and I've been fixing their drivers for past 6 months almost every day basis.)

Geeee you must have bad luck, I have never had a problem !! Then again I usually only install the driver and not catalyst control center.

 

[TheMailMan78]

The stupid thing here is that if you choose to "opt-out" and disable the limiter, you'll most likely void the card's warranty...

Who cares. You do that anyway the second you do any OC.

 

[Regeneration]

Some more slides are available here.