Looks like an accurate slide for 6970/Cayman specs

[Gloomy]

Oh, my bad. Here's the imageshack ul.
http://img809.imageshack.us/img809/6274/6970.jpg

I guess people take photographs with their calculator in China.

 

[busydude]

I guess people take photographs with their calculator in China.

Maybe its a 3D projector.. LOL.

 

[flopper]

http://www.frazpc.pl/artykuly/652601,Architektura-ukladow-Radeon-HD-6970-i-Radeon-HD-6950.html

Polish site if you use translator.

 

[mwmorph]

Rich people don't drive a Ford unless it is a Ford GT. If you are rich, you don't have to tell people you bought a fully loaded car. The car speaks for itself!

Well you also remember, rice people tend to also buy cars with option packages that exist. For example Supercharger+dealer install+12 month warranty+automatic does no exist anywhere in Ford's policies.

The particular situation he claimed is actually uniquely stated as not possible according to Ford for the first bullet point in product features and is also written in all caps to state the importance.

 

[Silverforce11]

I guess people take photographs with their calculator in China.

Probably one of those tiny pinhole cameras. Clip on tie etc. Its a confidential brief, you think they let you wip out cameras?

 

[MrK6]

I'm liking the 96 texture units. This card looks like it could be a beast at 2560x1600. I hope the NDA does lift tomorrow .

 

[RussianSensation]

I'm liking the 96 texture units. This card looks like it could be a beast at 2560x1600. I hope the NDA does lift tomorrow .

They have all the key ingredients in place:

2GBs of Ram
AMD's superior 8AA efficiency
AMD's superior texture fill-rate advantage (which the 5870 already had in spades)

Without a doubt, it will be a beast for that resolution.

 

[RaistlinZ]

Dumb question, but for the 6990, will all 4GB of RAM really be useable, or are we limited to the 2GB of RAM each GPU individually controls/has access to?

 

[Daedalus685]

Dumb question, but for the 6990, will all 4GB of RAM really be useable, or are we limited to the 2GB of RAM each GPU individually controls/has access to?

2GB per GPU, both GPUs need local access to the data and thus far there is no way to share a large pool without latency issues. (It works more like Raid 1 instead of Raid 0)

As far as I know (I'm only extensively familiar with the 4xxx series crossfire) the memory on each card (or GPU in a dual product) is simply an exact clone of the other so you can only load up to 2GB of unique data to the memory.

 

[MrK6]

They have all the key ingredients in place:

2GBs of Ram
AMD's superior 8AA efficiency
AMD's superior texture fill-rate advantage (which the 5870 already had in spades)

Without a doubt, it will be a beast for that resolution.

I'm still wary, tbh. The GTX 580 didn't set the bar too high imo and if that's all AMD is shooting for it still might not be that impressive. The main thing I'm holding out for is overclockability. If these things can't overclock for crap then they're useless to me and I'll sit this round out. If overclocking is great, then there are two scenarios I'm looking at: if the 6970 can offer similar performance to the GTX 580 at a lower price tag, such as $400 and within 5-10%, or can offer much more performance, in the 15-20%+ range, for $500, I'll be interested.

 

[chewietobbacca]

With those specs... I'm not sure they're only shooting for the 580 now...

 

[Idontcare]

Without a doubt, it will be a beast for that resolution.

At some point, presumably, these things need to become beasts at the elevated multi-monitor resolutions as well.

GPU reviews at single-monitor resolutions are one thing, but when people are touting 3-way setups as the next uber-rig then the reviews ought to be sampling that end-user space as well IMO. (and not as an after-thought or "feature test")

 

[Arkadrel]

So? 512 cores is less than 1600 and yet the GTX580 still destroys the HD5870.
Numbers aren't everything, unless the number is FPS.

Shaders are not cores <.<

The 5870 has 1600shaders/5 = 320 "cores"

vs

480s haveing "480" cores , 580 haveing "512" cores.


This new 6970 is supposed to have 1920/4 = 480.

** edit: Barts has 224, which is less than half of the 480 number, so its very likely this card might endup twice as fast in some instances, if the above is true. This would mean that the 6970 would be a faster card than the 580, if we follow that logic (im not saying that its the truth or likely to end up that way, I have no idea).

Your wrong about nvidia haveing "less" but more effective cores... its the other way around.

 

[3DVagabond]

At some point, presumably, these things need to become beasts at the elevated multi-monitor resolutions as well.

GPU reviews at single-monitor resolutions are one thing, but when people are touting 3-way setups as the next uber-rig then the reviews ought to be sampling that end-user space as well IMO. (and not as an after-thought or "feature test")

Agreed, but until nVidia gets it to work with a single GPU we won't see it as a general feature of reviews.

 

[Arkadrel]

From the slide about the "power containment unit"

* Integrated power control processor, monitors power draw every clock cycle.
--* Dynamically adjusts clocks for various blocks (of the gpu) to enforce TPD.
* Provides direct control over GPU power draw (as oposed to indirect via clock/voltage tweaks).
*No longer needs to constrain clock speeds to allow for outlier applications.
*User controlable via AMD overdrive.


So this power control processor, is able to not just adjust the total GPU by voltage/clock speeds, but to power down certrain parts of a GPU not needed for certain things ect, to save on power.?

It might even be able to prioritise whats most needed, so it down clocks X unit, and then up clocks in speed Y unit thats more needed. This way you can stay within a certain TPD but get better performance for same TPD.

in short... it ll increase performance/watt.

My prediction, these 69xx cards will be more performance/watt than the older cards are.
I also think its very likely we ll see future new gen mobile cards feature this unit, it was probably developed with this in mind but is also added to desktop card series ( Im guessing AMD are trying to hold onto mobile market share).

 

[bryanW1995]

So he's rich why are you guys giving from flak ..

read the entire thread. he said that he bought an automatic mustang for 40k, but the nice go-fast parts to go with that he pictured were for a manual tranny only. either he lied or he's completely clueless and just get's daddy to buy him everything, but either way he dropped out of the thread once he realized they were on to him.

 

[chewietobbacca]

From the slide about the "power containment unit"

* Integrated power control processor, monitors power draw every clock cycle.
--* Dynamically adjusts clocks for various blocks (of the gpu) to enforce TPD.
* Provides direct control over GPU power draw (as oposed to indirect via clock/voltage tweaks).
*No longer needs to constrain clock speeds to allow for outlier applications.
*User controlable via AMD overdrive.


So this power control processor, is able to not just adjust the total GPU by voltage/clock speeds, but to power down certrain parts of a GPU not needed for certain things ect, to save on power.?

It might even be able to prioritise whats most needed, so it down clocks X unit, and then up clocks in speed Y unit thats more needed. This way you can stay within a certain TPD but get better performance for same TPD.

in short... it ll increase performance/watt.

My prediction, these 69xx cards will be more performance/watt than the older cards are.
I also think its very likely we ll see future new gen mobile cards feature this unit, it was probably developed with this in mind but is also added to desktop card series ( Im guessing AMD are trying to hold onto mobile market share).

What'll be interesting is if they do it the reverse direction... like it's Turbo Core for GPU's

Would be interesting to see what AMD can do with that

 

[Sylvanas]

From the slide about the "power containment unit"

* Integrated power control processor, monitors power draw every clock cycle.
--* Dynamically adjusts clocks for various blocks (of the gpu) to enforce TPD.
* Provides direct control over GPU power draw (as oposed to indirect via clock/voltage tweaks).
*No longer needs to constrain clock speeds to allow for outlier applications.
*User controlable via AMD overdrive.


So this power control processor, is able to not just adjust the total GPU by voltage/clock speeds, but to power down certrain parts of a GPU not needed for certain things ect, to save on power.?

It might even be able to prioritise whats most needed, so it down clocks X unit, and then up clocks in speed Y unit thats more needed. This way you can stay within a certain TPD but get better performance for same TPD.

in short... it ll increase performance/watt.

My prediction, these 69xx cards will be more performance/watt than the older cards are.
I also think its very likely we ll see future new gen mobile cards feature this unit, it was probably developed with this in mind but is also added to desktop card series ( Im guessing AMD are trying to hold onto mobile market share).

That's a pretty big step for AMD to make this generation. Powering down specific on chip units to remain within a given power envelope would be a big investment at the archietecture level. I wonder if next gen cards from GF will feature power gating like Intel uses currently?

 

[3DVagabond]

That's a pretty big step for AMD to make this generation. Powering down specific on chip units to remain within a given power envelope would be a big investment at the archietecture level. I wonder if next gen cards from GF will feature power gating like Intel uses currently?

I hope it works OK. they don't need another problem with power management like the 5800 series had in the beginning. This looks like it could even be more sensitive to OS updates, etc...

I know why they're doing it. It allows them to squeeze more performance from a particular power level, rather than having to reduce the performance for a card overall just because certain applications, like Furmark, would push the card beyond acceptable limits.

 

[AtenRa]

From the latest info, I estimate

CAYMAN XT

Dual Front End (Graphics Engines) with double Tessellator Units
30 SIMD 4-way VLIW
1920 Shaders (64 per SIMD) (900MHz = 3,45 Tflops)
120 TMUs (4 Tex Units per SIMD)
2GB 256-bit Memory
32 ROPS
Performance ~30&#37; avg more than HD5870 (DX-9/10)
Performance ~2x DX-11 Tessellation more than HD5870
Price ~$450


CAYMAN Pro

Dual Front End (Graphics Engines) with double Tessellator Units
28 SIMD 4-way VLIW
1792 Shaders (64 per SIMD) (775MHz = 2,77 Tflops)
112 TMUs (4 Tex Units per SIMD)
1GB 256-bit Memory
32 ROPS
Performance ~15% avg more than HD5870 (DX-9/10)
Performance ~1.8x DX-11 Tessellation more than HD5870
Price ~$350

 

[BathroomFeeling]

Are those prices real, or guesstimates?

 

[AtenRa]

Not real, guesstimates

 

[Arkadrel]

I know the leaked spec slides say only "more than 160 GB/s Memory bandwidth".

But Im pretty sure that if they got chips for 5.5 Ghz - 6.0 Ghz they did so for a reason. I believe the 6950 will have 176 GB/s and the 6970 will have 192 GB/s, by means of running the chips at the speed they are designed to run at.

256x1.500 / 8 x 4 = 192 GB/s (the 6970) (1.500 x4= 6.0Ghz)
256x1.375 / 8 x 4 = 176 GB/s (the 6950) (1.375 x4= 5.5Ghz)

I just dont see the 6970 comeing out with only 160 GB/s, not when it has so much more in terms of shader power, along with everything else.

I also believe the tessellation performance will be greater than 2x of the 5870s.... if by mearly useing 2 tessellation units alone you reach 2x, without factoring in anything else. I belive theyve made improvements in efficiency compaired to the 5870s and 6870's.

even the slides of the 68xx cards where mentioned has haveing 2x over the 5870s... (which they did under certrain factors of tessellation). Surely theyve done more, if alone by adding a extra tessellation unit? should we not be expecting 4x tessellation performance (if 5870->6870 was a 2x jump, and now there are 2 tessellation units in the 6970 vs 1 unit only in the 6870)?

 

[AtenRa]

Tessellation needs Shaders and Rasterizers, that&#8217;s why AMD made the 6900 series a dual core architecture, each Tessellator unit have a 15 SIMD array (960 Shaders) and one Rasterizer. We could say that 6970 is a dual 6800 in regards to Tessellation.

 

[chewietobbacca]

30&#37; in DX9/10 seems a bit conservative..

80TMU -> 120 TMU (50% more)
ROPS internally are 2x that of 5870
850 -> 900 MHz core clock
320 -> 480 SIMDs (50% more)
1GB -> 2GB RAM

Even ignoring tesselation, that's some serious improvements. So unless memory bandwidth is a serious bottleneck (wasn't the case w/ the 5870s), we should see quite an increase in DX9/10