AMD: R7xx Speculation

not everything out there is true Shaidar, bits and pieces are but not the whole pie.

 

Which bits and pieces?

 

512-bits aggregate for the X2 is the only reasonable configuration. Also I think it's reasonable to assume GDDR5 won't clock as high as 2.5GHz this early.

Jawed

 

Certainly. Having said that, please point out anything I have stated that is most definitely not true.

 

theoretically what you stated is right, but only in circumstances where the gpu will be limited serverally in all those categories and all those categories are going to happen

 

I don't perceive any implication on my behalf that performance will be 160% greater than RV670 on average, or even ever. I'm just talking specs. For all we know it could be 200% faster

 

When they quote 2Ghz GDDR5, doesn't that mean 4Ghz effective? So 128GB/s on a 256bit bus?

If the memory speed of 1985Mz is to be believed then that should offer ~127GB/s n a 256bit bus.

Or did I get that completely wrong?

 

You're right, but GDDR5 isn't ready yet, and RV770 is in production right now so we won't be seeing GDDR5 coupled with an R7xx GPU until later this year.

 

GDDR5 is not ready yet!

When I checked on net best I could find was everyplace was saying first half of 2008. Last I checked May seemed to be in first half of 2008.

Memory is never ready yet until the first product comes out that includes it.

Is it not possible that ** if ** the 4800 series is an anouncement with product available in June and it being the first use of GDDR5.

It would seem to me, GDDR4 was a failure for the memory manufacturers as it never took off and it would behoove them to get out a credible replacement asap.

 

Qimonda started sampling gddr5 in Nov (Hynix a couple weeks later), I would think they'd be able to ship at least some small quantity soonish.

 

True.

Its not as if they have to phase out volumes of Gddr4 production.

It seems to me the effort to manufacture for a product line (4800) that wants to use the memory exclusively, is a good thing. They would make the effort and as AlphaWolf says , they have had samples for testing for quite a while.

 

Assuming it has 32 TMUs and 480 shaders... in the R600 series, the texture units are grouped into four, and each one feeds a specific shader array (correct?). So without the 1-to-1 ratio, would that mean that they're grouped differently, or that each texture group is not tied to any particular array?

(Doesn't the Xbox360 use a crossbar to feed the shader arrays from the texture units?)

 

A another article with picture.

http://www.tgdaily.com/content/view/37093/135/

Getting close... I wonder how on target they getting for specs.

 

They may be grouped only a little differently, i.e. with 8 "TMU" per block (4 blocks total) instead of 4 like in the RV670, or, better, to have 2 "Quad TMU" per each SIMD "level" instead of one (in this case and with 480 "SP" there would be 6 SIMD instead of 4). I don't know if the transistor count will be really in the 800 Million transistor number or not, however...

 

A few more TMUs and SPs, a few tweaks here and there, and GDDR5, but from the looks of things, no big changes from the 2900.
It's what the 3000 series should have been.

 

I'm really thrilled with the 32 TMU news. It's interesting that the fill rate will only now be catching up to the lower high end of current gen of Nvidia cards, even while Nvidias NEXT gen is around the corner. In other words AMD is still behind the ball there, even after doubling, which just goes to show how amazingly far back they were in the first place. Still, it's a start.

At 800 million transistors, I dont expect this to compete with GT200's alleged 1.3 billion. I like the prices though, very affordable.

It sets up a few interesting scenarios as well. If GT200 is even a few weeks later than May, which seems all but a cinch, AMD is likely going to have the undisputed fastest card for a period of time. And that is likely to garner them some hefty sales from the enthusiast community in that period. The type of people that have to have the fastest all the time (and the low prices should make it a pretty easy sale). Further, if AMD is truly on a six month schedule, then the refresh might very well be hitting not long after GT200.

It seems like it might be an interesting case of diverging strategies, might it be valid to persue smaller less challenging chips at more regular intervals, than go whole hog with a monster chip strategy that may suffer delays like GT200?

Anyways, cant wait to see how next gen cards handle Crysis. Even a doubling of speed will leave that game still a tough challenge for GPU's. Considering with a 9600GT I get around 20-24 FPS with high settings at 1680X1050 and no AA. Even twice as fast will struggle to handle ultra high setting and AA. I suspect we wont truly see Crysis dominated for yet another generation!

 

I remember Kyle of Hardocp once saying the R700 was a multichip but not crossfire. Have there been any info on it?

 

I spoke to Richard Huddy about that last September. Obviously he wouldn't be drawn on the details, but the overall gist was that R700 multichip will have a few new multi-chip communication features in its bag of tricks, but it wouldn't be a solution that "appeared" as a single rendering device to the driver/api/application whatever.

R700 might not strictly be Crossfire but my impression is that it won't be a big divergence.

 

the picture is gone... is it available anywhere else?

thx

 

Hmm, there's now mention of Rv770 having a split clock domains ala G8x and G9x?

Considering that rumors now are that it has 480 SPs and 32 texture units on 16 ROPs, wouldn't a higher shader clock just keep the bottleneck in the texturing units?

Regards,
SB