AMD: R9xx Speculation

I would be happy for a die-shot this time. Cypress didn't show us in its [strike]naked[/strike] full glory.

Perhaps now that it is about to be replaced, AMD would be willing to release a die shot? Then again, I don't understand why they didn't do it in the first place. It's not as if NVIDIA hadn't jumped on the first opportunity to get an HD 5870 and thoroughly dissect it anyway…
 
What's there to learn in a die shot (for the competition that is) :?: Maybe I'm being dense and simplistic, but it seems like tech pr0n for the rest of us forum folk. :p
 
Perhaps now that it is about to be replaced, AMD would be willing to release a die shot? Then again, I don't understand why they didn't do it in the first place. It's not as if NVIDIA hadn't jumped on the first opportunity to get an HD 5870 and thoroughly dissect it anyway…
In fact , I would expect them to have a fully fledged morgue for ATi chips , including separate organs and limps (shaders and other logics ) , maybe even an anatomy textbook. :p
 
In fact , I would expect them to have a fully fledged morgue for ATi chips , including separate organs and limps (shaders and other logics ) , maybe even an anatomy textbook. :p

Stop it please!

I'm getting mental images of nV engineers getting caught red handed licking Cypress dies.
 
1920 SPs/ 4 wide VLIW = 480 TPs. 480 TPs/16 TPs per SIMD = 30 SIMDs. 30 SIMDs *4 TMUs per SIMD = 120 TMUs. How big would be this GPU? Bigger than R600?
I tried to make an approximate estimation based on RV770 die (I know it isn't very accurate, but...)

premise: RV870 (Cypress) with its 1600 SPs (5D ALUs), 80 TMUs, 256bit interface, 32 ROPs and front-end (+PCIe and the rest) has 330mm²

Bart (1280 SPs (5D ALUs), 64 TMUs, 256bit , 16 ROPs) ~ 269mm²
Bart (1280 SPs (4D ALUs), 80 TMUs, 256bit , 16 ROPs) ~ 277mm²
Cayman (1920 SPs (5D ALUs), 96 TMUs, 256bit , 32 ROPs) ~ 388mm²
Cayman (1920 SPs (4D ALUs), 120 TMUs, 256bit , 32 ROPs) ~ 400mm²

I made a calculation of RV840's die size, which resulted in ~ 178mm², so these numbers shouldn't be entirelly off. Anyway, they can be off in the case of more significant architectural changes...
 
Last edited by a moderator:
Anyway, they can be off in the case of more significant architectural changes...

How many "significant architectural changes" did G71 have compared to G70? Granted the first was on 90 while the latter on 110nm, but if you still compare transistor counts and die area there's something "off" without having any significant changes.

I'm concentrating on your 4D numbers mostly and I'd say you might be off by a small margin ;)
 
because of all the SP/TMU fun in this thread; here is my take:

900MHz; 96 TMU's; 384 x 4 SP's (=VEC4); 32 ROPs; 256bit x 1500MHz GDDR5

IMHO the 1920 SP rumor is wrong. Only the HD68xx behaves like it has 1920 SPs (VEC5)
 
20 isn't a power of 2.
Neither is 5. Or 48.

BTW:
„128” would match „4” and „64” and „1920” -. and would lead to 60 TMUs. Given AMDs focus on processing power, that might be an alternative.

Edit 2:
To be a little less unclear: 1.920 in 15 SIMDs with 32 4-way-processors and a Quad-TMU each. You'd need double the Wavefronts in flight though.
 
Last edited by a moderator:
If Charlie is right and the die size is 115% Cypress for 130% overall performance then it looks like the architecture is a winner.

It does look like 40nm problems still persist somewhat because judging from the leaked code-names each performance tier (chip) seems to have three variants. I guess they are planning ahead this time, so there ought to not be any red headed step children in the lineup like the 5830 was.

So if the above holds true, I wonder what kind of price scaling we're looking at?

$399for the 6870 with 2GB Ram?

$299 for the 6850

$249 for the 6830

$199 for the 6770

$149 for the 6750

???

I was speculating along the same lines. The lowest priced HD 5850 has fallen to $269 at newegg, with HD 5870 at $349. So it seems reasonable enough for the new parts to be priced at $299 and $399. Im betting both the parts will be equipped with 2GB memory standard

As for the 6770, i dont expect it to be priced greater than $169-179 with the 6750 at $139-149.

I m afraid it will be higher. If Cayman is bigger than Cypress, it means that less Cayman chips will be produced per-wafer, which means less profit for ATI. The GTX480 is the fastest card around and it's sold for 499$: if Cayman is faster, i would expect the same price. So at least 449$-499$. I guess the good old times of RV770 are over.

Well one could argue that when GTX260 was selling for $449, AMD priced the 4870 at $299 so i think its fair to assume the current pricepoints wont change much. As mentioned above, Im expecting the 6870 to be priced at $399

I originally wrote that, but then I thought "No, Dave wouldn't be THAT cruel."... In any case I don't think its the wafer price / yields which will dominate the price, I suspect it'll be the RAM for the highest end SKU and the fact that they don't have any competition at that price bracket. I guess introducing it at a higher price would mean they could clear out the older Evergreen GPUs without having to slash the prices too badly.

I guess they are going to be using some of the new 2gbit GDDR5 chips?
OTOH, shouldn't the supply constraints be coming down soon?

Well TSMC was supposed to double the supply of 40nm wafers in Q3 but then Nvidia has a lot more chips coming and AMD also has to allocate wafers to Ontario so the supply situation is unclear as of now.

2Gbit GDDR5 chips were also slated for introduction in Q3 so i'd expect Barts to stick with 128 bit so that it could use 4 2Gbit chips to achieve 1 GB. This would be the most cost effective. Juniper was stuck using 8 1Gbit memory chips which increased costs. Similarly Cayman could use the standard 8 memory chips on 256bit for 2GB VRAM
 
http://news.ati-forum.de/index.php/...termin-und-weitere-details-der-hd6000er-serie

So shall, according to an anonymous source, the introduction of the new graphics cards on time-series on October 12 this year, AMD plans. This year is no graphics card, however the upper fields first go to the starting line as this makes most times a pixel accelerator from the middle segment the beginning. Demnach kommt als erste der neuen Serie, die von uns bereits angekündigte Grafikkarte mit dem Codenamen " Barts ". Thus comes the first of the new series which we have already announced graphics card with the code name "Bart". AMD seems like the competition is no air to breathe give way, because just that day, also plans to present its middle-class nVIDIA graphics card GT430 on GF108 basis.

The following list you will see the new GPUs that graphics chip will replace the previous:

Caicos: Low-End-GPU ersetzt Cedar/RV810 Caicos: Low-end GPU replaced Cedar/RV810
Turks: Mid-Range-GPU ersetzt Redwood/RV830 Turks: Mid-range GPU replaced Redwood/RV830
Barts: High-Mid-Range GPU ersetzt Juniper/RV840 Barts: High-Mid-Range GPU replaced Juniper/RV840
Cayman: High-End-GPU ersetzt Cypress/RV870 Cayman: High-end GPU replaced Cypress/RV870
Antilles: Dual-GPU Top-Modell ersetzt Hemlock/R800 Antilles: Dual-GPU top-model replaces Hemlock/R800

Furthermore, we could bring in experience that the series HD6000er probably the first to CHL a digital PWM control of the company will have CHIL. Next is the fact that "Bart Pro" in an expensive design comes from the cooler and heat pipes have is about 2x8mm. In addition, "the Chamber Barts XT technology to use" Vapor make in order to keep a cool head.

In addition, there are improvements in the rendering of 3D Blu-Ray, the introduction of UVD3 come with.. Connections in the form of HDMI1.4 1.2, and DisplayPort are called.
 
Im sorry for the offtopic, im new, but what exactly is Dave Baumann's job at Amd? is he into marketing.. chip designing.. or something else?
 
Back
Top