AMD: R7xx Speculation

[Tchock]

It's 100% certain that GT200 is big.


It's almost certain that RV770 is comparatively small, and relative to the original "R700" ideal of mini-cores, very traditional.

If the RV670 was better balanced, it would have even less problems as minimum framerates won't be like what it is now. That's why IMHO its performance shouldn't be too indicative of the final experience index.

 

[Lukfi]

Now the question is, will RV770 be balanced better than RV670? There's been a lot of talk that too few texturing units hindered the R600's performance. Now, while all the unreliable sources I can think of claim the RV770 is going to have 32 TMUs and 480 SPs (which would mean a decrease in the ALU:TMU ratio), some people say the number of TMUs will stay at 16. In my lame opinion, this is quite illogical if shader count will increase and Z-ops performance will be improved as well, not to mention more memory bandwidth.

As to the dual-core nature of the R700, I seriously don't think that it will be anything different from R680, it will still essentially be a CrossFire on one card, perhaps just without the PLX bridge, but that's a minor detail.

 

[Berek]

It's 100% certain that GT200 is big.


It's almost certain that RV770 is comparatively small, and relative to the original "R700" ideal of mini-cores, very traditional.

If the RV670 was better balanced, it would have even less problems as minimum framerates won't be like what it is now. That's why IMHO its performance shouldn't be too indicative of the final experience index.

GT200 is a 65nm part, while the RV770 is a 55nm. Unfortunately that means hot for the GT200... RV770, if given enough price leeway, might have a few advantages under its sleeve.

 

[INKster]

GT200 is a 65nm part, while the RV770 is a 55nm. Unfortunately that means hot for the GT200... RV770, if given enough price leeway, might have a few advantages under its sleeve.

The main question remains: Is the GT200 really RV770's market segment counterpart ?
The G92b (55nm) looks more like a direct foe to AMD's new baby than GT200, judging from most information/rumor leaks.

 

[Lukfi]

INKster is correct, G92b is the RV770's competitor, R700 (X2) will go against GT200. But GT200 could very well be manufactured on 55nm half-node. If the chip was designed for 65nm from the start, doing an optical shrink wouldn't be a problem (the RV670 was originally a 65nm part as well). But, apparently TSMC is offering two variants of the 55nm process: the standard one, used by RV670 & friends, and the performance one, which will be used by RV770. And I'm not so sure whether 65nm chips can be shrunk to performance 55nm (I'm skeptical about it). Looks like RV770 has the upper hand here, but again, the increase in price might outweigh it.

 

[Karma]

INKster is correct, G92b is the RV770's competitor, R700 (X2) will go against GT200. But GT200 could very well be manufactured on 55nm half-node. If the chip was designed for 65nm from the start, doing an optical shrink wouldn't be a problem (the RV670 was originally a 65nm part as well). But, apparently TSMC is offering two variants of the 55nm process: the standard one, used by RV670 & friends, and the performance one, which will be used by RV770. And I'm not so sure whether 65nm chips can be shrunk to performance 55nm (I'm skeptical about it). Looks like RV770 has the upper hand here, but again, the increase in price might outweigh it.

Where did you hear about the two different 55nm process types, and how could they be different from one another?

 

[INKster]

Where did you hear about the two different 55nm process types, and how could they be different from one another?

Maybe from TSMC itself ?
Their 65nm process has even more than just two different main options.

 

[Lukfi]

=>Karma: A well informed source told me the RV770 will achieve higher clocks due to an improved (though still 55nm) manufacturing process. That probably means the 55GC and the link posted by INKster also answers my question about the possibility of shrinking 65nm chips to 55nm "performance".

 

[Karma]

Thanks! I didn't even know there was such a thing.

 

[ECH]

Speculating R700

If the R700 uses some sort of chip to chip interface to communicate between dies (similar to Rambus’ Flex IO) and offer shared memory would it be feasible to have a homogeneous multi-die solution? Or should a heterogeneous multi-die solution appear more feasible?

 

[compres]

I am wiling to believe, given the parallel nature of GPUs and all the time that has gone by with no new arch from ATI(basically since R600 came out), that R7XX might be a big surprise in terms of being multi chip that works as a single chip.

[/wishful thinking]

 

[Rangers]

It's 100% certain that GT200 is big.


It's almost certain that RV770 is comparatively small, and relative to the original "R700" ideal of mini-cores, very traditional.

If the RV670 was better balanced, it would have even less problems as minimum framerates won't be like what it is now. That's why IMHO its performance shouldn't be too indicative of the final experience index.

By who?

If GT200 is 1 billion trans, and RV770 is "800m+", that's not a huge difference.

Originally my assumption was GT200 was 1.3 billion, but the latest rumors say 1 billion.

 

[Lukfi]

=>ECH, compres: That would require implementing NUMA and perhaps some other kind of synchronization between the GPUs... not to mention the need for a fast interface between the two (>100 GB/s through PCB, is that even possible with a limited number of pins?)
Simply put, it's more feasible to do good ol' crossfire on a card than that. Why do you think they launched R680 in the first place? So they could solve the CrossFire driver issues before R700 comes.

 

[Tchock]

By who?

If GT200 is 1 billion trans, and RV770 is "800m+", that's not a huge difference.

Originally my assumption was GT200 was 1.3 billion, but the latest rumors say 1 billion.

Absolute size-wise, I mean. A 512-bit mem interface would bump up size considerably (Imagine if R600 was 256-bit, it would be pretty lean even on 80nm)

And ATI's R6XX cards do pack trannies more densely, not sure if it's by inherent design or they squeeze more by purpose. RV670- 666@190mm^2, even a shrunk 55mm^2 G94 would still be behind on transistor count, and I'm quite certain it's still bigger, though the difference could just be negligible.

 

[DegustatoR]

GT200 is a 65nm part, while the RV770 is a 55nm. Unfortunately that means hot for the GT200... RV770, if given enough price leeway, might have a few advantages under its sleeve.

RV670 was on 55nm and G92 was on 65nm but i can't say that G92 is hotter than RV670 -)
That being said, it's pointless to compare RV770 with G100 -- if you want to know what's hot you should compare RV770 X2 with one G100. And something tells me that X2 will be hotter...

 

[aaronspink]

=>ECH, compres: That would require implementing NUMA and perhaps some other kind of synchronization between the GPUs... not to mention the need for a fast interface between the two (>100 GB/s through PCB, is that even possible with a limited number of pins?)
Simply put, it's more feasible to do good ol' crossfire on a card than that. Why do you think they launched R680 in the first place? So they could solve the CrossFire driver issues before R700 comes.

You don't need 100 GB/s through the PCB. 40 GB/s should be plenty and that is doable today with bog standard PCI-E signaling across a PCB.

Aaron Spink
speaking for myself inc.

 

[Rangers]

Absolute size-wise, I mean. A 512-bit mem interface would bump up size considerably (Imagine if R600 was 256-bit, it would be pretty lean even on 80nm)

And ATI's R6XX cards do pack trannies more densely, not sure if it's by inherent design or they squeeze more by purpose. RV670- 666@190mm^2, even a shrunk 55mm^2 G94 would still be behind on transistor count, and I'm quite certain it's still bigger, though the difference could just be negligible.

Could just be different transistor counting methods.

 

[Tchock]

Could just be different transistor counting methods.

And apply that to the original RV770 "800m+" and GT200 quotes, add a big mem interface, and we're back.

 

[Lukfi]

Absolute size-wise, I mean. A 512-bit mem interface would bump up size considerably (Imagine if R600 was 256-bit, it would be pretty lean even on 80nm)

I don't think the ~35M transistors would really make such a difference, although with 256bit interface they would be able to simplify the PCB design.

And ATI's R6XX cards do pack trannies more densely, not sure if it's by inherent design or they squeeze more by purpose. RV670- 666@190mm^2, even a shrunk 55mm^2 G94 would still be behind on transistor count, and I'm quite certain it's still bigger, though the difference could just be negligible.

The G94 is approximately the same big. But nVidia can't shrink G94 to 55nm process, nor they can't pack the 65nm trannies more densely, because the die size is already about the lower limit for 256bit interface.

You don't need 100 GB/s through the PCB. 40 GB/s should be plenty and that is doable today with bog standard PCI-E signaling across a PCB.

That comes out to 80 lines of PCIe 2.0. I don't think that's feasible. Even if they managed to connect the two chips with a really, really fast bus, there would still be a bunch of other problems to solve.

 

[nicolasb]

Chip Hell: http://bbs.chiphell.com/viewthread.php?tid=21168&extra=page=1

Referenced by The Inq here: http://www.theinquirer.net/gb/inquirer/news/2008/04/28/rv770-xt-yielding-better

A FEW MURMURINGS out of China suggest that DAAMIT has had better-than-expected yields and performance out of its latest RV770 chips, which means that the XT variation - using GDDR5 memory - should be en route to hardware enthusiasts sooner than expected.

RV770 Pro should be hitting the shelves in June at the latest, and it was previously thought that the XT variation could be as much as two months behind it. However, the fact the ramp-up has exceeded AMD's expectations means that the XT should now follow no later than a month after the Pro in mass availability, and the two cards will almost certainly ship simultaneously to press for review.