AMD: R8xx Speculation

[psolord]

Rumour-ish: AMD DX11 Cypress is Radeon HD 5870 & HD 5850

So the 5850 will be based on Cypress after all and not on Juniper!?

As i said the other day, a 150-170euro 5850 with double the performance of the 4850, will bring loads of cash in AMD's pockets!

Oh please AMD, make it so that the 5870X2 will not be more than 450 euros!

 

[ferro]

"Feel the power of more than 2 teraFLOPS plowing through the most demanding games."

That would be (more than) 47% more flops than the HD4890, 67% more than the HD4870, and 100% more than the HD4850. Assuming that this is for a mainstream single GPU card, that would be very good. No idea yet how this translates into actual performance of course.

 

[no-X]

At the time of RV770, the "1 TFLOP" was related to HD4850. I'd expect the 2 TFLOP card will be based on the fastest GPU, but not on the top configuration.

There are two possibilities for a 2 TFLOP board: 1280SPs at 800MHz or 1600SPs at 650MHz.

 

[liolio]

Also have info about first LRB sampledates to partners, but that's for another topic I guess.

A little question, did you post those informations here? I didn't find them after a little search
(or you run a website? thus I would happily consult it as you seems to have reliable sources )

 

[Ailuros]

Just make it something slightly shorter than 2.2 TFLOPs and it should get a lot warmer IMO

 

[Kaotik]

So the 5850 will be based on Cypress after all and not on Juniper!?

As i said the other day, a 150-170euro 5850 with double the performance of the 4850, will bring loads of cash in AMD's pockets!

Oh please AMD, make it so that the 5870X2 will not be more than 450 euros!

Of course it's based on Cypress, anyone who actually thought 5850 and 5870 wouldn't share the same chip should seriously take a moment and think these things through once or twice more.

Juniper has always been known as the middlerange product, in other words, 5600-series.

 

[Vincent]

Some things VR-Zone didn't mention yet:

- display outputs: DVI/DVI/DP or DVI/HDMI/DP. ATi Eyefinity support as of Cat 9.10 (8.66). Converter is included.
- "Feel the power of more than 2 teraFLOPS plowing through the most demanding games."
- "Get up to 1.8X of a graphics performance boost with ATI CrossFireX™ technology in dual mode."

So basically all the branding info partners need to make their boxes and stuff.

Also have info about first LRB sampledates to partners, but that's for another topic I guess.

What about the memory speed difference between 5870 and 5850 ?

 

[leoneazzurro]

So, 2 teraflops can be achieved with:

1200 SP at 833 MHz
1600 SP at 625 MHz
2000 SP at 500 MHz

if the shader unit uses the same architecture of R6xx/R7xx. According to the rumored die size of Cypress, I think 1200 SP is a little too low for the shader count of the new single ATI GPU flagship.
As this should be "more than" 2 Teraflops then it could be that Cypress XT is something like 2000 SP @ 600-650 MHz (about 2.5 Teraflops) or 1600 SP @ 750 MHz (2.4 Teraflops). I would not be surprised if the first solution will be right, as the lower clock could also mean that faster versions could come later should the GT300 and derivatives be a lot faster at launch than Cypress.

 

[psolord]

Of course it's based on Cypress, anyone who actually thought 5850 and 5870 wouldn't share the same chip should seriously take a moment and think these things through once or twice more.

Juniper has always been known as the middlerange product, in other words, 5600-series.

Sorry i meant the Cedar chip, not the Juniper!

Your argument still stands though!

So the Cedar chip is going to be the 5770 or what?

 

[Vincent]

That would be (more than) 47% more flops than the HD4890, 67% more than the HD4870, and 100% more than the HD4850. Assuming that this is for a mainstream single GPU card, that would be very good. No idea yet how this translates into actual performance of course.

All I want is the double (4870) " MSAA 4 AA+ 16 AF " performance based on the current I7 platform.

 

[w0mbat]

Who said its exactly 2TFLOPs? Maybe HD5850 ist 1600SPa@675MHz and HD5870 is at 850MHz+.

 

[fellix]

I would bet for 1280SP SKU.

 

[rpg.314]

I agree, a 1280 alu cypress is more likely.

 

[trinibwoy]

Yeah language like that usually means it's between 2 and 2.5 TF. Which is a pretty wide range for all sorts of combinations of unit counts and clock speeds. 1280x850 would fit nicely, so would 1600x750.

 

[mboeller]

I agree, a 1280 alu cypress is more likely.

or maybe 640 double pumped shaders (= dual shaders)

According to a chiphell comment the RV870 reached 850MHz. So 16 clusters with 1280 shaders seems possible.

Also neliz posted @ tweakers.net that ~ 300000 cards will be available at launch.

 

[Arnold Beckenbauer]

I would bet for 1280SP SKU.

Without beefy SIMDs, plz.

 

[Psycho]

Considering it's all for the box printing it must apply to both, including the 5850 (could be 1600@650), which probably makes the '70 around 2.5.

 

[mczak]

with 1280SP all you need for 2TFlops (assuming r6/7xx style shaders of course, if it sticks to rv770 organization that would likely be 16 clusters, each cluster with 16x5 and quad-tmu) is 782Mhz. That should be easily doable, rv790 did more than that clock stock already, and rv740 could be very easily pushed to that. So if both cards should have more than 2TFlops maybe something like 785 Mhz for the Pro (remember 4850 is exactly 1TF and not one flop more) and 950 Mhz for the XT (this is same clock afaik as the highest clocked 4890 shipped with, and if this would be just a shrink should be possible) could work (that difference between the two sounds a bit small though).
The rumoured die size of 300mm² sounds a bit high for this though I guess maybe, like rv790, it would also include a decap ring (or other similar measures which increase die size but don't really do anything functionality-wise, "only" help to achieve higher clock frequencies).

 

[fellix]

...and it fits nice with 64 TMU config (finally catching G92!).

 

[neliz]

170% of RV770 sounds about right.. as a start. I'm guessing bandwidth limitations

and the 300k number comes from debunking that deathsayer over at XS (that mentioned 25k total including October production) since those predictions would give a yield rate way less than 10% for 40nm. if we stick around 40% and a very modest amount of TMC capacity we already encroach 100k in good dies. at 60% and aggressive ramps since june you can come up with at least 300k.