AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

I think CU count has to be a multiple of 4, so 10CU isn't possible—unless GCN is a bit more flexible than we were led to believe.
In addition, if Pitcairn has the same front-end as Tahiti (two setup engines), 24 CUs are more likely than 20. Since they're organised in groups of four, you need 4 or 6 groups of them (2 or 3 per setup) to maintain symmetry of the design.

My guess for Pitcairn would be
- same front-end,
- 24 CUs,
- 6 RenderBackends (24/96 color/Z-Stencil ROPs),
- 512 KB L2, 256 bit and some GPGPU-related stuff removed or at least reduced.

That would probably put it in the ~70% area ballpark, which would mean ~256mm² or pretty much the same size as RV770 and Barts, provided the 365mm² rumor for Tahiti is corect.
 
In addition, if Pitcairn has the same front-end as Tahiti (two setup engines), 24 CUs are more likely than 20. Since they're organised in groups of four, you need 4 or 6 groups of them (2 or 3 per setup) to maintain symmetry of the design.
No. Or how does the 7950 with 28CUs or 7 groups of 4 CUs work? A setup engine in Tahiti is probably not fixed to a subset of the CUs.
 
In addition, if Pitcairn has the same front-end as Tahiti (two setup engines), 24 CUs are more likely than 20. Since they're organised in groups of four, you need 4 or 6 groups of them (2 or 3 per setup) to maintain symmetry of the design.

My guess for Pitcairn would be
- same front-end,
- 24 CUs,
- 6 RenderBackends (24/96 color/Z-Stencil ROPs),
- 512 KB L2, 256 bit and some GPGPU-related stuff removed or at least reduced.

That would probably put it in the ~70% area ballpark, which would mean ~256mm² or pretty much the same size as RV770 and Barts, provided the 365mm² rumor for Tahiti is corect.

And Capo Verde:
- one setup engine
- 12 CUs
- 4 render-backends
- 512KB L2, 128bit
 
In addition, if Pitcairn has the same front-end as Tahiti (two setup engines), 24 CUs are more likely than 20. Since they're organised in groups of four, you need 4 or 6 groups of them (2 or 3 per setup) to maintain symmetry of the design.

My guess for Pitcairn would be
- same front-end,
- 24 CUs,
- 6 RenderBackends (24/96 color/Z-Stencil ROPs),
- 512 KB L2, 256 bit and some GPGPU-related stuff removed or at least reduced.

That would probably put it in the ~70% area ballpark, which would mean ~256mm² or pretty much the same size as RV770 and Barts, provided the 365mm² rumor for Tahiti is corect.

And Capo Verde:
- one setup engine
- 12 CUs
- 4 render-backends
- 512KB L2, 128bit


I took a bit of flak when I posted that back in July...
Though my theorectical Cape Verde was 16CUs.

LordEC911 said:
Originally posted July 24th 2011

My speculation:

79x0 (faster than GTX580)
3.7-4b trannies
280-320mm2 (less than 350mm2 IMO)
32CUs=2048SPs @ 900mhz = 3.7TFlops
128TMUs
32-48ROPs
256bit[2GB] @ 6.8ghz = 218Gbps (384bit[1.5-3GB] @ 5.4ghz = 260Gbps or @ 4.8ghz = 230Gbps)

78x0 (a slightly faster Cayman, equaling or slightly beating a GTX580)
~3b trannies
210-250mm2
24CUs=1536SPs @ 950-1000mhz = 2.9-3.1TFlops
96TMUs
32ROPs
256bit[1GB & 2GB] @ 5.8-6.2ghz = 186-198Gbps

77x0 (offers ~6950/5870/6870OC performance)
1.8-2.2b trannies
140-170mm2
16CUs
64TMUs
http://semiaccurate.com/forums/showpost.php?p=127762&postcount=46
http://forum.beyond3d.com/showthread.php?p=1569288#post1569288

Everybody was harping about how AMD has been doing half the highend to the midrange and the exception was Barts.
 
Last edited by a moderator:
ASUS 7970 696€

desktop_2011_12_17_19pp2ny.png


http://www.salland.eu/product/1082713/asus-radeon-hd-7970-3072-mb-hd7970-3gd5.html

:LOL: Cool. :cool:
Any volunteers to buy it? :devilish:
 
The max resolution in that listing isn't close to being correct.

Neither is DX version nor amount of DVIs, nor the clocks according to slides anyway (as 1GHz was advertised as where the chips (in general) can be OC'd to)
 
You can charge anything you want for a product you don't have. Nearly 20% of that is tax anyway. If it's faster than the gtx580 I wouldn't expect a lower MSRP.
 
You can charge anything you want for a product you don't have. Nearly 20% of that is tax anyway. If it's faster than the gtx580 I wouldn't expect a lower MSRP.

They are gonna shoot themselves in the foot, the leg, the arse and the head if that's the actual price. Only the super wealthy/fanboi's will spend £500 to upgrade 20% from a 6970 or a 580. Especially when the new NV card is expected only 2 or so months away. I've held out for 3.5 years for a card thats gonna be me beneficial increases over my 4870x2, at a decent price point. The 7970 is so far NOT this card.
 
They are gonna shoot themselves in the foot, the leg, the arse and the head if that's the actual price. Only the super wealthy/fanboi's will spend £500 to upgrade 20% from a 6970 or a 580. Especially when the new NV card is expected only 2 or so months away. I've held out for 3.5 years for a card thats gonna be me beneficial increases over my 4870x2, at a decent price point. The 7970 is so far NOT this card.


What if it's not 20%? I've heard that "The card will leave no room for questioning anything no matter which camp you vow for", if you get what I mean
Also, highend Kepler isn't really expected 'till almost 6 months from now
 
What if it's not 20%? I've heard that "The card will leave no room for questioning anything no matter which camp you vow for", if you get what I mean
Also, highend Kepler isn't really expected 'till almost 6 months from now

I *really* hope you're right, cos I'm itching to replace this card now, but I'm seeing specs that are falling fast (eg ROPs 64 -> 32, BW 8GHz -> 5.5GHz etc). I know that this is all rumour and nothing is confirmed, but it was looking like a 100% improvement max over a 6970 with the best of specs, and that has reduced in pretty much every way by now.

A £500 card better perform near to a 6990, or it's gonna be a laughing stock.
 
What if it's not 20%? I've heard that "The card will leave no room for questioning anything no matter which camp you vow for", if you get what I mean
Also, highend Kepler isn't really expected 'till almost 6 months from now

If it's consistently 20% faster than the GTX 580, what does that leave to question?
 
Back
Top