AMD: R7xx Speculation

[turtle]

Anyone read chinese? The OP from this article was on chiphell saying something about R700 being delayed to Q3 and pointed to this article he wrote:

http://www.pczilla.net/post/236.html

Not sure if it's of any importance, but thought I'd throw it out there.

 

[Shtal]

That's assuming that a) RV770 will be released before G100 and b) that those leaked specs are true...

I do believe that AMD "ATI" is NOT sitting down and wasting time

 

[Unknown Soldier]

R700 generation consists of five chips

The lowest end one is codenamed RV710 and this one should replace Radeon 3450 and 2400 series, powered by RV615 and RV620 GPUs.

It is followed by RV730 and RV740 and these two are set to replace RV630 and RV635 mainstream chips.

The RV770 is the ultimate replacement for RV670, the hearth of Radeon 3870 cards, while the R700XT will consist of two RV770.

Read More: http://www.fudzilla.com/index.php?option=com_content&task=view&id=6017&Itemid=1

 

[CJ]

The rumors posted by Fudzilla were already known for a while. Those 4 GPUs could've been found in several drivers for a while now. What's more interesting is this blurp at Chiphell http://bbs.chiphell.com/viewthread.php?tid=17259&extra=page=1&page=1 which says something about RV770 being released in May, dismissing all those 480 SP rumors as BS and saying that RV670 is an enhanced RV770. And someone also mentioned that die-size is around 250mm2 compared to 194mm2 for the RV670. So we can safely assume they added hardware AA resolve this time around if it doesn't have more SPs and maybe some more TMUs?

I love those rumors.

 

[Mariner]

So we can safely assume they added hardware AA resolve this time around if it doesn't have more SPs and maybe some more TMUs?

I love those rumors.

AARGH!!!!

No - we're not expecting hardware AA resolve. The perceived below-par AA performance of the R600 is nothing (or very little) to do with this shader resolve as discussed ad infinitum already. In fact, I think many have postulated that poor AF performance is more of a problem.

More TMUs would certainly therefore help improve performance as that is one area ATI's chips are greatly underspecified in comparison to NVidia's.

 

[ShaidarHaran]

AARGH!!!!

No - we're not expecting hardware AA resolve. The perceived below-par AA performance of the R600 is nothing (or very little) to do with this shader resolve as discussed ad infinitum already. In fact, I think many have postulated that poor AF performance is more of a problem.

More TMUs would certainly therefore help improve performance as that is one area ATI's chips are greatly underspecified in comparison to NVidia's.

I think beefed-up ROPs with higher stencil/Z rates and higher AA sampling rates (4x), in addition to more TMUs is what R6xx really needs. That and higher clocks of course.

 

[nicolasb]

For what it's worth;

http://www.fudzilla.com/index.php?option=com_content&task=view&id=6004&Itemid=1

We have seen some marketing talk from the DAAMIT machine and it was funny to read that its upcoming RV770 mainstream chip will end up twice as fast as RV670.

Well this is probably right but this would naturally be the best case scenario. You should count on average 20 to 30 percent performance improvement but in some resolution and at some settings, the new RV770 might really leave the RV670 aka Radeon 38x0 cards in the dust and score twice as fast.

It is possible as we could see with 9600GT that ends up twice faster than 8600GT but mainly because of the 128 versus 256bit memory controller and the fact that the 8600GT was barely faster than the old 7600GT to begin with.

 

[wishiknew]

All I want from ATI is just to make a bandwidth limited chip. You done the fancy memory stuff and lots of shader units. How about just brute force lots of tmus.

 

[no-X]

G92 has four-times as many TMUs as RV670 and is only about 30-50% faster

 

[ShaidarHaran]

G92 has four-times as many TMUs as RV670 and is only about 30-50% faster

That's because it's (G92) largely limited by its setup capabilities and memory bandwidth in most scenarios. Find a test-case that is primarily texture-bound and it's not unreasonable to expect G92 to be head-and-shoulders above RV670.

 

[no-X]

I know. And that's it. TMUs themselves aren't holy grail of 3D acceleration.

 

[trinibwoy]

I know. And that's it. TMUs themselves aren't holy grail of 3D acceleration.

Well neither is bandwidth or ALU's as evidenced by R6xx's relatively lackluster performance even when endowed with an abundance of both. I do find it curious that R6xx also has a pretty substantial advantage in pure geometry tests but still falls behind in geometry shader benchmarks. What happens at the front-end of the pipeline in these architectures is still a bit of a mystery to me.

 

[ShaidarHaran]

Well neither is bandwidth or ALU's as evidenced by R6xx's relatively lackluster performance even when endowed with an abundance of both. I do find it curious that R6xx also has a pretty substantial advantage in pure geometry tests but still falls behind in geometry shader benchmarks. What happens at the front-end of the pipeline in these architectures is still a bit of a mystery to me.

Depends on the GS test as well. Seems like the more complexity, the better R6xx does. At least in Shadermark. Could this be due to R6xx's superior branching performance/capabilities?

 

[AlexV]

Well neither is bandwidth or ALU's as evidenced by R6xx's relatively lackluster performance even when endowed with an abundance of both. I do find it curious that R6xx also has a pretty substantial advantage in pure geometry tests but still falls behind in geometry shader benchmarks. What happens at the front-end of the pipeline in these architectures is still a bit of a mystery to me.

It does extremely well in geometry tests due to the fact that vertex shaders map wonderfully onto it's SPs(there was a mention by Eric, I think, saying that this was one of the reasons for the way they setup the alus), so it's probably one of the cases where you get close to maximal utilisation of the SPs.

About the geometry shader benchmarks...dunno, I've seen only Rightmark2.0 Alpha/Beta done by the Digit-Life boys, and the way those tests are setup you're likely to be bound by other aspects of the chip(s) before actually hitting a GS bound so I wouldn't consider that a very indicative test.

 

[ECH]

Rumor Mill
Do you agree that if the next gen ATI video card (RV770) has 640 SP and 32 TMUs that it's a beast? In other words a video card that is not of an X2 variant (using AFR) using double the TMUs and SPs. Or do you still believe more rops are needed?

Source

The core frequency of RV770 is a little higher than RV670, but not over 1GHZ. The ram used by RV770 is GDDR5, and the core size is 250 mm2. The stepping A11 GPU's yield is over 71% when recently taped out. And the stepping provided to retail will be the same A11 stepping. Whether stepping A12 will be provided is not known yet. The detail specification of RV770 is just more than two times SP number and TMU number of RV670. The release date of RV770 will be some day in May.

 

[Pete]

OMG sounds like RV770's packing 1 extreme gigahurtz.

Fudz reads like its talking about a 1GHz RV670 with "tweaks" (though adding TMUs or zixel power would seem to suggest "super-linear" gains vs. RV670's clocks) when it says +20-30%, though it's hard to speculate when in the same breath it suggests bandwidth is the major reason why the 9600GT embarasses the 8600GT/S.

 

[turtle]

It's been flying around that forum the idea Rv770 is like how R580 was to R520, hence the die size only increases a small amount while still on 55nm, perhaps mainly increased because of the TMUs?

Hmm....so 10 MADD units per shader instead of 5 or 4 flops per shader unit instead of 2?

 

[Ailuros]

If the die size truly is around 250mm2, it could suggest something in between 175-190M more transistors. Something tells me that not all unit amounts have been doubled as some of the funky rumours suggest so far. It's enough if throughput is doubled, while not necessarily having in some spots twice as much units and it makes way more sense too. Combined probably with some further algorithmic optimisations and an expectable frequency increase, the result should have more than good chances to distance itself from a 8800Ultra f.e.

If all that is true it doesn't come as a surprise at all that NV is shrinking G92 down to 55nm. Even if they're preparing a single high end monster for the second half of this year, it'll take them time to get to a cost-effective Performance GPU out of that future family.

 

[Twinkie]

Just how big is the performance gap between a 8800Ultra to a HD3870 let alone a 8800GTS to a HD3870?

 

[Pete]

Just how big is the performance gap between a 8800Ultra to a HD3870 let alone a 8800GTS to a HD3870?

http://www.behardware.com/articles/701-14/product-review-the-amd-radeon-hd-3870-x2.html
http://www.computerbase.de/artikel/...e_9600_gt_sli/24/#abschnitt_performancerating