TheInquirer on R520 delay

[AthlonXP1800]

Looks like the inquirer is finally catching on.. and owning up to it actually being a 16 fragment design.

http://www.theinquirer.net/?article=24698

if 700MHz 16 pipeline design turn to be true then Nvidia already have the answer to it, it called G71. :D

[-=DVS=-]

They sure makeing alot of guesses :D

[shabby]

Looks like the inquirer is finally catching on.. and owning up to it actually being a 16 fragment design.

http://www.theinquirer.net/?article=24698

16 extreme pipelines? Heh

[BrianG]

We previously suggested that R520 will end up with 24 or even more pipelines but we are advised that companies are counting pipelines differently and that now it's all programmable.
Back-pedal. First it was 32 all the way, then 24 but the 32 pipe would be the flagship, then 24 would be fast at the high clocks and now 16.

[mattyo]

Back-pedal. First it was 32 all the way, then 24 but the 32 pipe would be the flagship, then 24 would be fast at the high clocks and now 16.
Damage Control :angel:

[coldpower27]

That would be interesting though, if the rumors are true then R520 with 16 fragment pipelines, it should have a good die size advantage against the 7800 GTX.

[-=DVS=-]

Was hopeing for something cool like 32(shader pipes?)/16rop @500mhz like 7800gtx is 24/16rops @430mhz , but oh well 16 rops at 700mhz would give huge fillrate numbers.

[AthlonXP1800]

Was hopeing for something cool like 32(shader pipes?)/16rop @500mhz like 7800gtx is 24/16rops @430mhz , but oh well 16 rops at 700mhz would give huge fillrate numbers.

Well R520 will hardly give huge fillrate numbers over G70:

R520 at 700MHz/16PP = 11,200 MPixels
Reference 7800GTX at 430/470MHz/24PP = 10,320/11,280 MPixels
BFG 7800GTX at 470/510MHz/24PP = 11,280/12,240 MPixels
XFX 7800GTX EE at 490/530MHz/24PP = 11,760/12,720 MPixels

[-=DVS=-]

Hmm i might be off , but doesn't 7800Gtx got different pipes not your conventional like 6800Ultra and X800 ?

7800Gtx - 24 but only 16 ROPs
6800Gt/Ultra/X800 - 16 rops
So if R520 got 16 Rops but at 700Mhz wouldn't that be completely different balance compared to 7800Gtx. This is a bit confuseing , need to read up on technology :D

And that 40+ Geometry clock doesn't add anything to Fillrate..its just internal thing, was already said by Nvidia employ. Shoud be ignored and only main clock used.
P.S BFG runs at 460mhz :p

[ChrisRay]

Your right. In single texturing bilinear filtered scenerios. The 16 pipe 650 mhz card would be much faster. The question you gotta ask yourself is. Is that important? Since shader routines and even multitexturing take several cycles anyway. The difference wouldnt be noticable in real world situations.

IMO the real question and differences might be seen in the AA since it takes several cycles through Z. Keep mind Nvidia does 32:0 Z stencils still even with 24 shader pipelines. Another thing to keep in mind is the shader per/pipe speed of the G70 verses R520. Which will have more efficient ALUS?

[-=DVS=-]

So what you saying is ATI should have went with similar design like Nvidia maybe up it to 32/16 at decent clocks ! instead of smaller 16 rops only design.

Xbox 360 gpu will be somewhat similar minus unifed architecture no ? only 8 rops was it and flexible 48 pipes made of 3x16 something (crazy).
Id imagine 32/8 rops and 16 for vertex work :confused:

[AthlonXP1800]

And that 40+ Geometry clock doesn't add anything to Fillrate..its just internal thing, was already said by Nvidia employ. Shoud be ignored and only main clock used.
P.S BFG runs at 460mhz :p

Opp my mistake, yes BFG run at 460MHz.

Well someone over at XtremeSystems (http://www.xtremesystems.org/forums/showthread.php?t=68146) forum overclocked ASUS Geforce 7800 GTX to 770/810MHz, 3D Mark 05 ran with the geometry clock, it recorded the score with 810MHz geometry clock, the fillrate should go up.

[ChrisRay]

So what you saying is ATI should have went with similar design like Nvidia maybe up it to 32/16 at decent clocks ! instead of smaller 16 rops only design.

Xbox 360 gpu will be somewhat similar minus unifed architecture no ? only 8 rops was it and flexible 48 pipes made of 3x16 something (crazy).
Id imagine 32/8 rops and 16 for vertex work :confused:


I didnt say that at all! :eek:

[DMA]

http://vr-zone.com/?i=2487&s=1

We heard that the last R520 spin at TSMC is at 3% yield rate for 24 pipes and we can certainly expect a much lower yield for 32-pipes R520

Not looking good if this is correct.

I'm tired of waiting so 7800 GTX here i come. :)

[mattyo]

http://vr-zone.com/?i=2487&s=1



Not looking good if this is correct.

I'm tired of waiting so 7800 GTX here i come. :)
They were wrong with the info on G71, wonder if we can still believe them. :angel:

[Graphicmaniac]

If ATI will sell a just 16 pipe card i think nvidia will start say they use "old" 16 pipe tecnology :D lol

[rohit]

why 16 extreme pipe, they can come out with a 20pipe card too. against 7800GT.

[MUYA]

why 16 extreme pipe, they can come out with a 20pipe card too. against 7800GT.
why a 20 pipeline? That could mean a difference i yeild...in terms of actual core speeds...what are the tradeoffs? will it be good enough process? etc

It really makes sense from a transistor to core clock ratio standpoint to have a 16 fragment pipeline GPU at 90nm but with high core clocks.

16 fragment pipelines...and at great core speed. That will equal or be greater than the fillrate of a 24 fragment pipeline card at lower core clock. Its a valid strategy. When you increase the amount of fragment pipelines you are not only increasing the level of transistor budgetry complexity but also driver complexity. Did you know that on the 6800U some fragment pipeline can also sit idle due some sort of scheduing conflict? Imagine that on a gpu twice the amount of fragment pipelines??

Indications are from a few sources that 32 fragment pipeline r520 on 90nm is highly unlikely....on paper

R520 on paper is every bit as competitive

[rohit]

LOL it was just an assupmtion abt. 20pipe thing.
16 fragment pipelines...and at great core speed. That will equal or be greater than the fillrate of a 24 fragment pipeline card at lower core clock. Its a valid strategy. When you increase the amount of fragment pipelines you are not only increasing the level of transistor budgetry complexity but also driver complexity.
i dont get this, considering-->
nv40 (16pipe) is 400mhz on core, r420 (16pipe) is 500+ mhz on core. both 130nm.
Then too both perform equally.. cometitively (almost).

[MUYA]

LOL it was just an assupmtion abt. 20pipe thing.

i dont get this, considering-->
nv40 (16pipe) is 400mhz on core, r420 (16pipe) is 500+ mhz on core. both 130nm.
Then too both perform equally.. cometitively (almost).
u are comparing apples to oranges in old to new gpu tech. Ppl should really get out of the "more fragment pipelin is better" metality. Everything is not dependant on it...there are many variablities.

[rohit]

hmm.. alrite.

[Ninjaman09]

u are comparing apples to oranges in old to new gpu tech. Ppl should really get the "more fragment pipelin is better" metality. Everything is not dependant on it...there are many variablities.
Wise words. :D

[AthlonXP1800]

They were wrong with the info on G71, wonder if we can still believe them. :angel:

Well actually they got most things correct, maybe they are correct that G71 are based on 90nm, it still unclear what process G71 are based on either 110nm or 90nm, it best to wait and see. :D

I guess 3% Yield on 24 pipelines are very bad and maybe 1.5% Yield on 32 pipelines must be very badly.

[mattyo]

Well actually they got most things correct, maybe they are correct that G71 are based on 90nm, it still unclear what process G71 are based on either 110nm or 90nm, it best to wait and see. :D

Fair then.


I guess 3% Yield on 24 pipelines are very bad and maybe 1.5% Yield on 32 pipelines must be very badly.
hehe, that is probably the worst rumor in the recent months surpassing all the stupid Inquirer ones. :lol:

[nutball]

I guess 3% Yield on 24 pipelines are very bad and maybe 1.5% Yield on 32 pipelines must be very badly.

The 24-/32-pipe yields could be even worse than that, they could be zero ;) What are the yields on 16 pipe dies that clock at the clock that they need to clock at?