ATI Radeon HD 4000 Series discussion

[AliG]

well, yes and no, g92 did fine with 16 ROPs, but g80 beat it once you cranked up the max details and resolution, that's why the 8800ultra beats even the 9800gx2 in some cases with extreme detail etc

[LordEC911]

meh im not sure about that, just look at 8800ultra vs 9800gtx -> in high res the ultra "stomps" the 9800gtx.

Very different, too many variables to make an estimation.
Bandwidth, frame buffer, TMU&TFU and ROPs.

[Helmore]

And that all has to do with the extra memory bandwidth the 8800 Ultra has.

[Hornet331]

Very different, too many variables to make an estimation.
Bandwidth, frame buffer, TMU&TFU and ROPs.

even without AA 8800ultra beats 9800gtx in everything higher then 1600x1200

9800gtx has higher shader clock, higher core clock, more TMUs same TAUs but lessbandwidth...


well even compared to the 8800gtx it loses sometimes in 1920x1200 and they have nearly the same bandwidth.

Im still concerned that 16 rops are not enough for the RV770 to do well in high resolutions.

[Helmore]

Im still concerned that 16 rops are not enough for the RV770 to do well in high resolutions.

I'd go as as to say that more than 16 ROPs is overkill for RV770.

[Macadamia]

even without AA 8800ultra beats 9800gtx in everything higher then 1600x1200

9800gtx has higher shader clock, higher core clock, more TMUs same TAUs but lessbandwidth...


well even compared to the 8800gtx it loses sometimes in 1920x1200 and they have nearly the same bandwidth.

Im still concerned that 16 rops are not enough for the RV770 to do well in high resolutions.

It's enough if they improved the Z and color fillrate capabilities.

RV670 has proven to be very flxible in Crossfire, reaching 2560 where the Ultra can't. It's pretty much obvious that if you want 2560 speeds, you have to pay $440 or $500 (4850 CF or 4870X2)

[STaRGaZeR]

I'd go as as to say that more than 16 ROPs is overkill for RV770.

I have the same opinion, since RV670 do almost "nothing" in the ROPs and RV770 is supposed to be a derivative of it.

[Nedjo]

I have the same opinion, since RV670 do almost "nothing" in the ROPs and RV770 is supposed to be a derivative of it.

Pfff… you really don’t have a single clue what are ROP’s (or in ATI’s semantics - render backend) used for, and generally about anything slightest technical about CPU or GPU architecture?!

[zerazax]

If they're still using shaders for AA then increasing RBE's won't matter that much (besides, they supposedly doubled the Z-rate)

And if you look at RV670, it seems to scale well with resolution, unlike the G92s which tank under higher res and settings

[AliG]

yes and but since it failed with detail, its still unclear whether or not 32 TMUs is enough

[cadaveca]

32 is enough. 24 is enough, possibly, too, if AA worked right, and not in shaders. r670 only works @ high res with no AA...even low res w/AA sucks.

[AliG]

No its about the lack of texture power, doesn't have much to do with AA not being done in the shaders. With double the texture power and double the Z rate, hopefully it should be enough for modern games with max detail. But something tells me if the 800 number is true, we'll be seeing 40 TMUs rather than 32 (same 4:1 ratio)

[Helmore]

32 is enough. 24 is enough, possibly, too, if AA worked right, and not in shaders. r670 only works @ high res with no AA...even low res w/AA sucks.

I would not be surprised if a future NVIDIA architecture did AA in the shaders as well. I'm basically expecting it to happen, it's more a matter of 'when' with NVIDIA. There is nothing bad about AA in the shaders, the ATI chips just lacked in Z-fillrate and lacked in texture filtering and other AF related things.
Actually AF gave the biggest hit when turned on on ATI's chips, AA was actually pretty good for the amount of Z-fillrate as it actually scales better than NVIDIA's AA. (for the amount of Z-fillrate)

[cadaveca]

No its about the lack of texture power, doesn't have much to do with AA not being done in the shaders. With double the texture power and double the Z rate, hopefully it should be enough for modern games with max detail. But something tells me if the 800 number is true, we'll be seeing 40 TMUs rather than 32 (same 4:1 ratio)

LoL. You mis-understand, really. TMU's are TEXTURE MEMORY UNITS...cache...

The real problem with R670 is AA implementation. If they fix that, they can get away with 24 TMUs, although this is truly dependant on how many shaders there are...the number of shaders will dictate the number of TMUs.

Helmore, while I partially agree with you, it does not see mto make much sense that AF was a big problem. 16xAF @ 2560x1600 is not a show-stopper. AA WILL make an app unplayable, so while texture filtering and z-ops were a bit anemic, they weren't the real problem.

AA in shaders is essentailly the future, yes. When we get 4xMSAA standard...but that's still a couple of gens off yet...more likely to appear the same time as raytracing...

So ATI had a foray into it, didn't work well for them, but i think they understand the problems they are facing. Doesn't mean they have to beat the dead horse...

[Rammsteiner]

I just read out of boringness some wiki stuff about Ati, and it says 4870 X2 should be able to do Octa CF?

Is that true?

[Helmore]

I just read out of boringness some wiki stuff about Ati, and it says 4870 X2 should be able to do Octa CF?

Is that true?

First a question for you:
How reliable is Wikipedia?

[bowman]

First a question for you:
How reliable is Wikipedia?

As reliable as its sources.

People usually won't add that for the hell of it, usually they read it on some site, mostly they don't bother adding source cites so it could be from anywhere from anandtech to fudzilla.

If one wonders, might as well just google it to find the original news post.

[Xello]

http://en.wikipedia.org/wiki/Radeon_R700

Editors at least follow wiki guidelines and post sources, just check them at the bottom. Mostly Fudzilla and Inq.

With the severely diminishing returns of going from 3 gpu crossfire to 4 gpu, can you really see them doing 8 I guess that dude who rigged up 4 98 GX2's could do it.

[STaRGaZeR]

Pfff… you really don’t have a single clue what are ROP’s (or in ATI’s semantics - render backend) used for, and generally about anything slightest technical about CPU or GPU architecture?!

Nice try. And yes to the question. I'll talk seriously when I have the card in my hands.

And I'm still waiting your 3DMark06 3870CF 2,4GHz vs 3,0GHz numbers

[madcho]

The octa CF is possible.

the CF of the R6xx can work with 2^n graphicards. ( the Trio of asus is not faster than a X2 ).

8 works 16 works, ... and go on !!!

[AliG]

LoL. You mis-understand, really. TMU's are TEXTURE MEMORY UNITS...cache...

The real problem with R670 is AA implementation. If they fix that, they can get away with 24 TMUs, although this is truly dependant on how many shaders there are...the number of shaders will dictate the number of TMUs.

Helmore, while I partially agree with you, it does not see mto make much sense that AF was a big problem. 16xAF @ 2560x1600 is not a show-stopper. AA WILL make an app unplayable, so while texture filtering and z-ops were a bit anemic, they weren't the real problem.

AA in shaders is essentailly the future, yes. When we get 4xMSAA standard...but that's still a couple of gens off yet...more likely to appear the same time as raytracing...

So ATI had a foray into it, didn't work well for them, but i think they understand the problems they are facing. Doesn't mean they have to beat the dead horse...

Whatever you want to call it, they need a higher texture fill rate, put it that way

[cadaveca]

For what? Halo? Doom2?

Seriously tho, under PS2.0, ok. nVidia excels in this situation most often, but thier shader domain is twice as fast as ATI's, and hence the difference. TMUs aren't the issue...clockspeed is.

By your means, they would have to double TMUs to 32 to supply thier 320 shaders with the needed fillrate, and if the talk of 800 shaders is correct, then we'd be seeing 80TMU's, not the 40 you suspect.

Texturing is only important as much as you think in older apps, not in SM3.0...and both nV and AMD cards favor math-intensive processing for texture heavy-processing...and when it comes to lighting, RV670 excels, especially with phong.

Apps with TWIMTBP are coded to take advantage of nvidia's texturing power...they are old apps, after all. There's nothing wrong with rv670 and texture fetches...they just need to be twice as fast, not necessarily twice as many!

[AliG]

No that's not completely true, if that were the case, then ATI cards wouldn't die under any detail, especially AF. They do fine in pure resolution and that's why we know that its not a lack of shader power (though the clockspeed is why they aren't as powerful as nvidia's offerings in pure resolution)


And also, in order to double the texture fill rate, we're talking either double the core clock, which won't and will not ever happen, or doubling the count, therefore doubling the count wins

[cadaveca]

Um, TMU's do the texture processing? NO!!! The shaders do! So yes, that IS the case! More specifically, the inconnection speed between the TMUs and the shader core is much faster in nV's solution, hence the difference.

But when you add in math-intensive texture operations, like PHONG lighting, ATI excels, due to the pure shader count, and the efficiency of the dispatch processor. THis is why the increase of TMUs does not make a ton of sense, as you surmise, as there's more at play here than just the number of units, but also thier speed, and how they interconnect with the gpu, as well as thier size and precision!

[Dan_c]

what do you guys think the r700 will do in Open GL?!
vs. g92 I mean... still don't have a clue