AMD RV770 refresh -> RV790

Not really arguing with that, but don't forget there are other potential issues why 4830 and 4850 could perform similar, one being that it could be setup limited.

Yes, I'm always puzzled about setup limitation. I have one assumption though - it surely can't be general? I mean, while Crysis with its 2-3M tris / screen is surely susceptible to such a limitation, a game with average 0.6-0.8M tris are probably not or much less so. Which means that the two cards should get closer and closer as the triangle count increases.

Though I agree that 8 RBEs would quite likely limit performance at least to some degree.
Seems to me the rumors aren't quite converging yet - if it should achieve HD4830-like performance (though model number doesn't exactly suggest this) then 640SP, 32TMU, 16 RBEs (and 128bit, 900Mhz GDDR5) sounds logical. But no way that's going to fit into a 100mm² die.

My "ideal" rv740 would be 480 ALU, 24 TMU, 8 RBE and 128-bit gDDR5, clocked at 750/1000. That would trail the 4830 by some 3-15%, and would also fit in 100mm2. But probably AMD has difference perf targets.

@Arnold: that would be a bit of a shock, wouldn't it? :LOL: anyway, it seems to me that AMD is deliberately avoiding the use of the term "ROP", that's why I tried to do that, too.

@Jawed: thanks for the analysis!
 
Didn't you guys discuss the question what die size would be necessary for a 128 bit memory interface just the other day? And was 100mm² the magical number? :)
 
My "ideal" rv740 would be 480 ALU, 24 TMU, 8 RBE and 128-bit gDDR5, clocked at 750/1000. That would trail the 4830 by some 3-15%, and would also fit in 100mm2. But probably AMD has difference perf targets.

As an IHV you do not want cards like a HD4830. At some point your yields will be so good that you have to start crippling good GPUs that could have been sold as HD4850s just to meet demand.

It happened in the past with R580 and X1900GT. ATI replaced the crippled X1900XT(X) as fast as they could with RV570 which was a native 36 ALU card with 12 ROPs and a 256-bit memory interface. And look how successful the X1950Pro turned out to be.

So it seems logical for them to do the same again with the popular HD4830... phase it out as fast as you can in favour of a native solution (albeit this time with a 128-bit memory interface and GDDR5 to offset the difference).
 
As an IHV you do not want cards like a HD4830...

Yeah I know harvesting, and I fully agree with everything you wrote - my only question is, is it inherently necessary that the replacement is faster than the old card? And, will the 640 ALU/32 TMU/2 RBE-quad variant at 700MHz be really faster than the 4830?
 
Yeah I know harvesting, and I fully agree with everything you wrote - my only question is, is it inherently necessary that the replacement is faster than the old card? And, will the 640 ALU/32 TMU/2 RBE-quad variant at 700MHz be really faster than the 4830?

Other examples are the 9600XT over the 9500Pro. It doesn't necessarily have to be (that much) faster to be a good replacement.
 
I have to admit I think it's very likely: 8 RBEs and 800MHz GDDR5's bandwidth (according to VR-Zone), 51.2GB/s, is actually pretty much on target for that - it's the kind of bandwidth RV730 should have had.

There'll be trouble though if it doesn't reach at least 750MHz.

Jawed
 
I wonder whether or not all the rumors floating around about green editions of 9800 GT/GTS240 and 9600 GT (with no new name :( ) have something to do with 40nm-problems.

Anyone else got the feeling, Nvidia will be trying to emphasize lower power consumption (albeit also lower performance FWIW) as long as they do not have their own 40nm-GPUs ready? I mean, with a new name and DX11 not immediately coming out, plus no need for additional power connectors - maybe opposed to RV740, !!!dunno, rumors!!! - they surely could score some OEM design wins for back to school season, right?

The green stuff is an obvious marketing angle(ie change the criteria by which the card is evaluated in the consumer mind). Dont think the green editions have anything to do wih 40nm. Pretty sure its more an inventory thing. A disguised way you can drop the product into the larger market segment below to clear it out quicker.

If you look at what they are doing the 9600s they were trying to sell before standalone with the higher power requirements can now be sold in oem systems much more easily as they wont emit nearly as much heat, likely make as much noise or stress the low capacity power supply. Also tells you which cards nvidia have too much of though....

Dont think the GT218 will be that far behind the RV740 once 40nm is working. Whether GT216 will be ready for back to school is a question. The margins on the GPUs have dropped quite a bit lately, enough that it might be in nvidia's interests to concentrate more on IGPs as the volumes are better.

Back on topic - Tchock's ATI lineup didnt have a GDDR3 RV790 Pro part. Samsung recently announced they've doubled the efficiency producing GDDR5 parts and are expecting it to be 50% of high end gpu market by years end. If ATI is also planning a GDDR5 RV740, then will need quite a large chunk of that capacity wont they?
 
If I'm not interpreting CJ's tidbits wrong that signature might be by multitude of lanes longer than reality.

***edit: die size doesn't fit that picture either by the way.
 
Last edited by a moderator:
I suppose I should add that if RV730's "refresh" is effectively pad-limited by its 128-bit bus, so it gets a doubling in ALUs, then why shouldn't RV790 get something similar?

RV730 is operating in a more price sensitive part of the market - and currently losing to 9600GT it seems. And 9600GT won't be its competition for ever, either. AMD has no choice about using a 128-bit bus, so the question is, how best to fill out the spare die space with clusters/RBEs.

I'm not convinced it'll be competitive with 9800GT though - which is why I'm dubious about it being only 100mm2 with the 8 clusters, 640 ALUs 8 RBEs configuration. At least this configuration has enough TUs :LOL:

It does seem strange to me that some rumours said that RV740 was due to launch before Christmas - within 3 or 4 months of RV730 seems pretty weird.

Jawed
 
RV730 is operating in a more price sensitive part of the market - and currently losing to 9600GT it seems.
Speaking of price sensitivity: Don't forget that 9600 GT is about 20 percent more expensive (at least here in germany).
 
Does the current RV770 use the TSMC 55nm GP process? If so could this RV790 be just a move to 55nm GT process to increase the clocks along with some other architectural "refinements"?
 
Back
Top