AMD Cayman info (or rumor)

[Loque]

amd should let some benchmark leaks happen to prevent people from buying the gtx580.. unless it isn't as fast :p

[LightSpeed]

http://www.fudzilla.com/graphics/ite...-november-29th @

FUD, right? I hope...

Hes right actually...

edit: I mean about the launch date which has been pushed back to the 29th, might even be early december for decent availability

[Ourasi]

@Loque: Nah, they want to hurt the 580's value as bought today so their owners will whine like crazy in all nV threads around and give them all the free publicity they could ever wish for...


Seriously, the handfull of guys buying this two weeks before their launch have enough doe to spend and would not care to much about losing some of it, or would buy nV anyways to get the glorious FisX-Cooda shait....

[Macadamia]

Chiphell says there's a Radeon press event in Beijing on the 22nd.

Could be like the Eyefinity event briefing like the 5800s, but more on performance instead.

[zerazax]

Chiphell says there's a Radeon press event in Beijing on the 22nd.

Could be like the Eyefinity event briefing like the 5800s, but more on performance instead.

Not sure about the translation. One also reads that the conference is this week, release 22nd

It appears Cayman XT might be arranged in a possibly 6 x 5 x 16 x 4 SP modules with PRO aranged in 6 x 4

[spursindonesia]

Some tidbits from a certain forum that name shall not be mentioned.

I'm pleased to post some rumours from three graphics discussion boards in the Mainland China.

Firstly, according to the three sources (Chiphell's nApoleon, someone in ZOL and Edison in PCinLife),
Cayman XT is a 1920 SP chip.

Secondly, according to someone in ZOL,
Cayman XT consists of six modules, each with one tessellators.

Thirdly, according to the one who posted OpenCL Benchmark Results of the 1536 SP sample card on Chiphell,
TMUs no longer attach to the SIMD Cores. 4 (x4) TMUs are shared in the module.

nApoleon also mentioned about a conference in Mainland China on 22 November.

THIS is getting very interesting.

[Oztopher]

I get the first point (1920 SP's), but i'm not in tune with the other two, what do they mean exactly?

[spursindonesia]

I get the first point (1920 SP's), but i'm not in tune with the other two, what do they mean exactly?

Some sort of optimization of AMD's massive VLIW design meets Fermi's distributed geometry processing ? Optimise your own advantage and take others good idea & adapt it for your own design ? Perhaps we will be dealing with a kinda freakenstein designed chip.

We all know nVidia's Fermi has its own merits and advantages in its mArch design when it comes to DX11 gaming, while IMHO AMD mostly only used legacy design with DX11 flavor implemented in it (quite cheaply) for its Evergreen line up, making it more easy to produce & on time to market.

[zerazax]

I get the first point (1920 SP's), but i'm not in tune with the other two, what do they mean exactly?

So currently Barts is arranged in 2 rows of 7 SIMDs - each SIMD has 16 x 5 = 80 SP's

Likewise, Cypress XT is arranged in 2 rows of 10 SIMDs - again, each having 16 x 5 = 80 SP's

It appears that Cayman may very well be arranged in 6 modules - each with 1 row of 5 SIMDs - each SIMD is 16 x 4 = 64 SP's, thus giving each module 320 SP's and thus an overall setup of 1920SP's. Cayman PRO may thus have 1 SIMD in each module disabled, thus creating 6 x 4 SIMDs for 1536 SP's total

And for the third point, it appears that TMUs are no longer tied to each row of SIMD, and thus they can be shared - that is, disabling SIMDs doesn't disable TMUs.

More interestingly, it appears that Tesselation will be tied to each module, and going from 2 engines to 6 should mean a significant boost in performance

[Oztopher]

I see, thanks, guys

Cayman looks to be very well on track then. Whatever happens, both line ups will be very close. I guess it will come down to certain things like power consumption, overclocking etc.

Edit: AMD should bring out a 6930 as well, to hold its own against the GTX 560, as i don't think the 6870 will do very well against it... maybe level in DX10 performance and less demanding DX11 games, but it will get thrashed if it doesn't have caymans suped tessellation and geometry processing.

Then barts can go head to head with the GTS 550 (if there is such), just like the 5750/5770 vs GTS 450 currently.

[SKYMTL]

What do you get after reading this?

http://www.bit-tech.net/blog/2010/11...ch.net+feed%29

It means that press were briefed about certain aspects of the Cayman cards a while ago. I thought that was quite obvious to many when the HD 6800 series reviews were launched.

[Iconyu]

It means that press were briefed about certain aspects of the Cayman cards a while ago. I thought that was quite obvious to many when the HD 6800 series reviews were launched.

Yeah I read one review where there made the mistake of saying the 6870 were 4D, so I assumed he'd mixed up his material from the Cayman briefings.

[marten_larsson]

http://www.fudzilla.com/graphics/ite...-november-29th @

FUD, right? I hope...

I really like the part about AMD having more cards than the competition expected - Fud getting his FUD from Nvidia?

If the rumours from China are correct it could be a true monster performer...

[SKYMTL]

Yeah I read one review where there made the mistake of saying the 6870 were 4D, so I assumed he'd mixed up his material from the Cayman briefings.

I find it interesting that people hold up that "4D" indicator without even knowing what it means.....

[Iconyu]

I find it interesting that people hold up that "4D" indicator without even knowing what it means.....

They switched from 1 complex + 4 simple to 1 complex + 3 simple to increase the ratio of complex shaders right?

[marten_larsson]

They switched from 1 complex + 4 simple to 1 complex + 3 simple to increase the ratio of complex shaders right?

From what I've gathered they've changed from 1 complex + 4 simple to 4 semi-complex. Some of the instructions use more than one shader for the most complex stuff but some stuff that could only be done by the complex unit is now possible in all 4 new. Gipsel over at Beyond3d together with Jawed over there has been discussing it quite a bit.

[hatter]

They switched from 1 complex + 4 simple to 1 complex + 3 simple to increase the ratio of complex shaders right?

No, according to rumours it means they are going for 4 shaders with medium complexity instead of 1 complex+4 simple config

[ajaidev]

From what I've gathered they've changed from 1 complex + 4 simple to 4 semi-complex. Some of the instructions use more than one shader for the most complex stuff but some stuff that could only be done by the complex unit is now possible in all 4 new. Gipsel over at Beyond3d together with Jawed over there has been discussing it quite a bit.

+1 that is what i also gather 4D is...

[RaZz!]

Waiting is hardly ever a good idea, there's always something better coming. The only time you could've gone wrong with early adoption is with the GTX280/260.

Before and after that, it was generally the rule: buy now, be happy later. (R300, G80, RV800)

hum, in general, you're right - BUT since cayman already knocks at the door and will be released in the next few weeks, you'd be dumb to not wait, imo

even if you hate ati/amd or whatever, maybe cayman will force nvidia to lower prices on their gtx580. so it's a win-win for you, the consumer.

[Sly Fox]

even if you hate ati/amd or whatever, maybe cayman will force nvidia to lower prices on their gtx580. so it's a win-win for you, the consumer.

My thinking as well. Competition is good no matter which side, if any, you're rooting for.

[Iconyu]

From what I've gathered they've changed from 1 complex + 4 simple to 4 semi-complex. Some of the instructions use more than one shader for the most complex stuff but some stuff that could only be done by the complex unit is now possible in all 4 new. Gipsel over at Beyond3d together with Jawed over there has been discussing it quite a bit.

Seems like the most flexible method, you could turn the entire core over to complex if needed. But when I first heard of it, it sounded like too much like a pipe dream. Assuming 1920 shaders, 4+1 would allow for 384 vliw at any one time, 3+1 gives 480 vliw and 4 mid-complex would give 960.

I don't know how much work per clock you can get in comparison to nvidia's double clocked shaders but it seems to me there would definitely more opportunity to maximise that proposed AMD set up in comparison to nvidia's.

[Your_Boss]

Cayman Pro 6950 - 3 rows, 8 SIMD, x 16 x 4 = 1536
Cayman XT 6970 - 3 rows, 10 SIMD, x 16 x 4 = 1920

OR

Cayman Pro 6950 - 6 rows, 4 SIMD, x 16 x 4 = 1536
Cayman XT 6970 - 6 rows, 5 SIMD, x 16 x 4 = 1920

Sounds about right.

[SimBy]

What do you get after reading this?

http://www.bit-tech.net/blog/2010/11...ch.net+feed%29

What I find funny is how everyone is amazed at 20% performance improvement over GTX480. How the hell can that be impressive?!

I still remember the bashing AMD got when HD5000 series was not DOUBLE performance over HD4000 series.

Srsly WTF?

[MrMojoZ]

What I find funny is how everyone is amazed at 20% performance improvement over GTX480. How the hell can that be impressive?!

I still remember the bashing AMD got when HD5000 series was not DOUBLE performance over HD4000 series.

Srsly WTF?

I was impressed that they spent the effort to address the heat/noise etc instead of just forging ahead with another half baked design. And they priced it realistically. Nvidia has pushed me away over the past few years so I'm hoping this is the start of a new trend.

[SpuTnicK]

Found this:

Dont know if this is trustworthy,

Source