AMD: R7xx Speculation

[Arun]

Look at the hardware.fr numbers; they only test in two configs: 16xAF and 16xAF+4xAA. It's pretty clear that G94 doesn't just have a substantial advantage with AF; it also has one with AA.

It's also pretty clear that G94 becomes less attractive at higher resolutions, which is to be expected as framebuffer compression becomes easier (->G9x's advantages there matter less) and the average number of texture cycles per TEX operation becomes lower, thus increasing the effective ALU-TEX ratio.

 

[Arun]

It's the same wall actually =)

No it's not, because going multi-chip isn't going to help you at all for the power wall. It doesn't matter whether you go single-PCB or dual-PCB; you just can't cool a dual-slot solution that dissipates 400-500W! I'm not even sure watercooling could save you there...

 

[DegustatoR]

No it's not, because going multi-chip isn't going to help you at all for the power wall. It doesn't matter whether you go single-PCB or dual-PCB; you just can't cool a dual-slot solution that dissipates 400-500W! I'm not even sure watercooling could save you there...

Maybe you weren't listening -- i said that you'll need _multiple_ coolers once we hit the wall. Each GPU in this multi-GPU solution needs it's own cooling system. So it's basically a multi-card or even multi-system solution, were each card has not only it's own cooler but it's own power supply too.
I know it sounds awful, i'm just thinking out loud =)

 

[AlexV]

No it's not, because going multi-chip isn't going to help you at all for the power wall. It doesn't matter whether you go single-PCB or dual-PCB; you just can't cool a dual-slot solution that dissipates 400-500W! I'm not even sure watercooling could save you there...

Not wieldy watercooling, in any case....meaning not something that an AIB could ship in a box and price competitively(I'm thinking about the entire watercooling enchilada, not only the waterblock...because if you include only the waterblock you get guys building a loop with a single 120 Rad and a weakling pump and wondering why they have trouble with thermals).

 

[mczak]

Look at the hardware.fr numbers; they only test in two configs: 16xAF and 16xAF+4xAA. It's pretty clear that G94 doesn't just have a substantial advantage with AF; it also has one with AA.

That's not quite universally though, in some benchmarks the HD3850/3870 cards get even faster with AA enabled compared to the competition - in Oblivion actually by a significant margin (I wonder what's up with that), in Crysis (but only high/DX10!) not much, in others they at least don't get slower (Colin McRae).
But yes, in some other benchmarks rv670 definitely loses more ground with 4xAA (even more compared to 9600GT than 8800GT, obviously).
hardware.fr benchmarks with TRAA though, seems reasonable but it could have some impact, I'd like to see numbers with "pure" MSAA.

 

[Jawed]

Now what 'competition' are you talking about? AFAIK NV was R&Ding multi-GPU solutions far longer than ATI/AMD.

Ignoring Rage Fury MAXX, R300 is able to work in multi-GPU systems:

http://forum.beyond3d.com/showthread.php?t=2680

see last paragraph.

The screen-space tiling paradigm introduced with R300 is still going strong in R6xx and NVidia took up the idea for G80 it seems.

Jawed

 

[Ailuros]

I think what people kinda mean with the whole multi-core mumbo jumbo in GPUs is that it's easy for the hardware designers to scale it. A CPU like Barcelona that's designed for 4 cores with a shared L3 cache should be incredibly easy to scale to 8 cores as long as the cache & memory controllers have been designed with an arbitrary number of cores in mind.

I don't think many GPUs in the industry's history were like that. Many derivatives either weren't on the same process node or had changes here or there. The whole notion of a "family" in the GPU world is more about the (majority of the) RTL than the synthesis or the verification phases. This also leads to necessary differences in the drivers. There might be a few exceptions, but that's not the point.

Certainly you might be able to improve this and reduce your per-chip costs that way, making it more attractive to go single-chip at the high-end and also release more low-end derivatives to target more price points 'natively'. It's not a proven strategy, but I guess it could work. Is that what you're thinking of, Ailuros? Also, regarding a transistor count wall: that's senseless, we'll hit a power wall way before then, most likely.

More or less yes; and by the way I consider that transistor count/power wall somewhat imaginary; as long as there will be effective alternative sollutions I don't see any IHV running against a brickwall in the longrun. We had in the past neverending myths about imaginary bandwidth walls; where the hell are those anyway?

 

[Ailuros]

It's not just STALKER either which exhibits a large hit when only 16xAF (but not AA) is applied. There are now lots of reviews of the 9600GT, and I had to read countless times how the 9600GT has a much smaller AA hit without any evidence whatsoever that it wasn't in fact a much smaller AF hit (argh) (though, arguably, the 9600GT should indeed have a very small AA hit, smaller than 8800GT).
Some numbers from the computerbase 9600GT review (http://www.computerbase.de/artikel/hardware/grafikkarten/2008/test_nvidia_geforce_9600_gt_sli/) when only 16xAF was enabled but the game didn't support AA, this should even be without HQAF btw:
Gothic 3 1600x1200 no AF: 9600GT 15% faster than 3850 512MB
Gothic 3 1600x1200 16xAF: 9600GT 63% faster than 3850 512MB
(the 9600GT has pretty much free 16xAF here, the 3850 certainly has not...)
Stalker 1600x1200 no AF: 9600GT 24% faster than 3850 512MB
Stalker 1600x1200 16xAF: 9600GT 52% faster than 3850 512MB
UT3 1600x1200 no AF: 9600GT 4% slower than 3850 512MB
UT3 1600x1200 16xAF: 9600GT 4% faster than 3850 512MB
(not a huge difference this time but enough to swap rankings of the cards)
Bioshock 1600x1200 no AF: 9600GT 2% faster than 3850 512MB
Bioshock 1600x1200 16xAF: 9600GT 5% faster than 3850 512MB
(not really a difference here, this game does not appear to be texture limited at all since both cards essentially get 16xAF for free)

So if the rv770 indeed somehow has more texture units this should help it being more competitive in some games rv670 currently isn't really (at least not with the quality options AA/AF enabled). If that happens, I already bet that everybody will write "AA is fixed!" when in reality it's just fast as ever but the AF hit got smaller... That's not to say though the performance hit from enabling AA might not get smaller indeed, if ROPs get updated or memory bandwidth increased.

I don't think RV670 is limited in terms of bandwidth at all, especially if one compares all its fillrates to bandwidth ratios compared to G92.

I won't say that future Radeons won't benefit from more texel fillrate, yet it'll equally benefit from a higher pixel/Z-fillrate as higher arithmetic throughput. Have a look also at the 3800X2 results in those benchmarks; with 16xAF in Stalker f.e. it's exactly twice as fast as a 3870 and that with 26.4 GTexels/s, which might end up not being a couple of notches lower than the fillrate of a RV770.

Let's skip for a moment the above rather outdated game engines and let's have a look at resource hog Crysis under "high" settings:

http://www.computerbase.de/artikel/...idia_geforce_9600_gt_sli/21/#abschnitt_crysis

Look at the performance drop of the 9600GT in 1280 from 1xAA/1xAF to 4xAA/16xAF compared to a 8800GTX or Ultra for instance. It's probably a result of better compression and/or better z-culling, improvements I'd also suspect in RV670 compared to R600.

Anyway I'd find it very ironic if in the foreseeable future RV770 should lower its ALU:TEX ratio compared to predecessors while NVIDIA might increase its own on contrary.

 

[Arun]

I moved all the posts regarding AMD's balance sheet and operating expenses to "The AMD Execution Gloom Thread", as suggested by BRiT. Please keep this thread focused on R7xx and, to a lesser extend, multi-chip solutions in general. Thanks!

 

[Shtal]

I was thinking lately about ATI RV770, GF8800Ultra and (G9x) GF9800GTX.

Looks to me until Nvidia releases G100, ATI/AMD would be able to compete again with single GPU Ultra_High-End Market - as example back in the days of R580 vs. G71.

 

[DegustatoR]

Looks to me until Nvidia releases G100, ATI/AMD would be able to compete again with single GPU Ultra_High-End Market - as example back in the days of R580 vs. G71.

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

 

[Kaotik]

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

Based on current rumors/info, RV770 should come close to 6 months before GT200/G100, as the former is expected late Q2, latter late this year

 

[Ailuros]

Based on current rumors/info, RV770 should come close to 6 months before GT200/G100, as the former is expected late Q2, latter late this year

That's only one quarter apart based on those very same rumours and I don't expect to see a RV770-X2 at the same timeframe as RV770 either.

 

[Kaotik]

That's only one quarter apart based on those very same rumours and I don't expect to see a RV770-X2 at the same timeframe as RV770 either.

Was it late Q3/early Q4 for the GT200/G100? I just remembered late this year which in my mind translated to late Q4

 

[Arun]

Uhm, GT200 is late Q2 or (I presume) more likely early/mid-Q3: http://www.fudzilla.com/index.php?option=com_content&task=view&id=5172&Itemid=34

 

[Ailuros]

Uhm, GT200 is late Q2 or (I presume) more likely early/mid-Q3: http://www.fudzilla.com/index.php?option=com_content&task=view&id=5172&Itemid=34

Let's give upcoming products a tad more pathetic lifetime *cough* (ducks under his desk) *ahem*. I hadn't read that one but I figure that if all goes according to plans yadda yadda for both that "R700" Fudo means might appear roughly around the same timeframe as NV's next big thingy.

Hit me with Wavey's frying pan, yet it's a pleasant surprise to see AMD execute that well for a second time

 

[Dave Baumann]

Assuming RV670 is the "first time" then I'm guessing you forgot RV620 and RV635.

 

[compres]

Assuming RV670 is the "first time" then I'm guessing you forgot RV620 and RV635.

I thik r300, r350 and r420 were on time

 

[Ailuros]

Assuming RV670 is the "first time" then I'm guessing you forgot RV620 and RV635.

Τhere aren't any virgins in this story obviously

 

[Squilliam]

With the speculation on the performance of this chip, isn't it of interest how WELL the g94 performs relative to it's g92 cousins. If we consider this effect and multiply it by being a larger/faster chip. Wouldn't it be conceiveable under these circumstances that it would be 50% faster as the leaked slides suggested?

In addition to this, could the performance disparity between the G92/G94 chips be due to the g92 being rushed to market to compete with the r670? If they have a respin for the 9xxx era they could be a much better competitor than we realise?