AMD: R9xx Speculation

Not sure this is worth posting, but pcper has written a short speculation article here. His take on the VLIW-4 arrangement seems particularly strange to me... isn't it the t-unit that is supposedly underutilized? Why go to 3+1?

I can't see anything else new there, except PCPerspective suggesting there will be no improvement in tesselation...but that could be based on the lack of information on the weeks-old slides from chiphell.
 
C
I wouldn't expect that. If you adjust for process node, NVidia has been getting worse per mm2 since G92. Well, maybe GF104 was a step up from GT200, but they're pretty close. IIRC, ATI took a step back with Evergreen (again, adjusting for process node).
GF108 also seems to do better in perf/area vs gt215. The ddr3 versions of the cards (gt430 vs. gt240) have pretty similar performance, and GF108 is like 20% "simpler" (both less transistors/smaller).
 
133526h71fh7zu4e17ijfp.jpg


http://www.chiphell.com/thread-129466-1-1.html

One question : why a rectangle , not a square ?
This will give us a much needed food for technical discussion , instead of meaningless arguments about nomenclatures !
 
One question : why a rectangle , not a square ?

Clusters of SIMDs arranged in 2 or 3 blocks versus a single stack as in the squarer Juniper & Rv770 die.

"Dual Cypress style engine" or whatever that slide said.

And not fully square like Cypress cuz there aren't that many of them to fill it out. I'd imagine Cayman squarer looking.


edit: Btw, thats 960 SP, 1280 would be 280ish range
 
Pricing will be key here; NI seem to match up so closely to their intended counterparts so AMD would do well to undercut them by a bit. The pin compatibility mentioned above will definitely help with this.

A Barts derivative would slot nicely into a highend mobile role; the 5xxx derivatives will do just fine for the lower end mobile parts since they're not making a process transition. This seems like a keep nVidia-at-bay / architectural pipe-cleaner refresh.
 
Last edited by a moderator:
Looks like AMD went all out on Cayman XT. Huge TDP. I'm quite sure performance will be awesome.

I wonder what the die size is.

And it seems Antilles is indeeed 2x Barts. Prolly Barts XT. Not sure they could keep 2x Caymans PRO below 300W.
 
Last edited by a moderator:
Barts pin-compatible with Cypress? c00l, looks like they are ready to price-war NV (tho they'll prefer no to )

who will prefer not to? Do you see NV cutting another 30% of their die size soon to make them competitive again?

I'm looking forward to at least another 6 months of steady AMD pricing. yay for ASP!
 
Then undervolt & speed bin them and release antilles with lower clocks than cayman

Question is how much faster than Cayman XT would Antilles (2x cherry picked low power whatever Cayman) be since they would both be around < 300W. I guess I expected Cayman XT to be around 250W max. Something doesn't add up.
 
Question is how much faster than Cayman XT would Antilles (2x cherry picked low power whatever Cayman) be since they would both be around < 300W. I guess I expected Cayman XT to be around 250W max. Something doesn't add up.

The only thing the slide indicates is that 225W < Cayman XT < 300W. It could very well be 230W.

Hell, it could even be 224W and the extra power might only be there for overclocking purposes.

I guess those big dies and high TDP figures are the result of TSMC canceling their 32nm process, it seems AMD just implemented their designs in 40nm without changing anything substantial.
 
Last edited by a moderator:
6Gbps = 192GB/s, or about 25% more than Cypress. Performance gain is hopefully in excess of that.

MBA? Manufactured board assembly?
 
Back
Top