ATI R600 Taped Out

[BassBomb]

it still baffles me people go to inq

 

[josh6079]

Originally posted by: BassBomb
it still baffles me people go to inq

QFT. Here's a big clue....look who wrote it:

Fuad was his name.....and look what it almost looks like.....

Faud+ Inquirer=Fraud

 

[BFG10K]

When we say big, it will be physically big

Uh-oh.

Does having a unified architecture mean that HDR and AA can be enabled always?

No.

Or another example soft shadows and AA in FEAR?

That's a game issue I believe.

 

[BenSkywalker]

Also, why do we keep increasing pixel shaders without bumping up the number of actual pipes on new cards? We've been stuck at 16 since the 6800/X800 days. What good is having 64 shaders if there are only 16 ROPS to feed them through?

Two parts- one is memory bandwidth. If you have 16ROPs then to write out that many pixels using a standard, last gen, 32bit color format you need to have 512bits of memory bandwidth per clock cycle on the GPU. That moves to 1024bits if you are utilizing FP16HDR(64bit color). Moving beyond 16ROPs would speed up certain functions, but until we see some eDRAM based solutions on the PC(too expensive) GPUs then we aren't going to see a puch to more ROPs anytime soon.

Second part- some shaders take hundreds of clock cycles to execute(some take MUCH longer then that) but developers can't use them heavily as nothing today has close to enough shader power. We could have 500 shader units for each ROP and it still would not be close to enough to handle CGI level shaders with today's level of complexity.

anyway, i'm of the opinion that the xb360 runs dx10, or something close enough to not matter, so that ati (which worked closely with MS to develop R500, i'm sure) knows what will be in dx10, so they can get a dx10 R600 out pretty easily

D3D10 offers very, very little in the way of new features. It just does things in a superior way as to not create so many bottlenecks. Don't expect much at all beyond D3D9 in terms of features- it is simply a different approach.

In the unified architecture, can more resources be dedicated to to say textures versus shaders and vice versa?? (ie. can it act like a 30 "traditional" pipeline part??)

No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs.

 

[thilanliyan]

Originally posted by: BenSkywalker

In the unified architecture, can more resources be dedicated to to say textures versus shaders and vice versa?? (ie. can it act like a 30 "traditional" pipeline part??)

No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs.

So what exactly is the benefit of the unified architecture??

 

[josh6079]

Originally posted by: thilan29

Originally posted by: BenSkywalker

In the unified architecture, can more resources be dedicated to to say textures versus shaders and vice versa?? (ie. can it act like a 30 "traditional" pipeline part??)

No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs.

So what exactly is the benefit of the unified architecture??

To decrease the amount of idle shaders and increase the amount of ones that are working. There are only a few instances on todays retail platforms where an application is drawing as much vertex shader operation as it is pixel shader. Because of this, there are very few instances where you are utilizing both aspects of a high end GPU. Since the shaders here are going to be unified, meaning that they can process both vertex and pixel, this will create a more efficient work load for a GPU since it isn't wasting architecture or its shaders.

 

[LittleNemoNES]

Originally posted by: josh6079

Originally posted by: thilan29

Originally posted by: BenSkywalker

In the unified architecture, can more resources be dedicated to to say textures versus shaders and vice versa?? (ie. can it act like a 30 "traditional" pipeline part??)

No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs.

So what exactly is the benefit of the unified architecture??

To decrease the amount of idle shaders and increase the amount of ones that are working. There are only a few instances on todays retail platforms where an application is drawing as much vertex shader operation as it is pixel shader. Because of this, there are very few instances where you are utilizing both aspects of a high end GPU. Since the shaders here are going to be unified, meaning that they can process both vertex and pixel, this will create a more efficient work load for a GPU since it isn't wasting architecture or its shaders.

Oh! I get it. Cool. Then these cards should be a big leap from current gen cards!

 

[fierydemise]

The best analogy I can think of for unified shaders (from my limited understanding) is a highway. Assume the highway has 8 lines, usually each direction only has 4 lanes no matter how much traffic there is in each direction. So it is very possible for there to be a traffic jam slowing things down in one direction and the other direction moves freely. With unified shaders the number of lanes in each direction can change depending on need, so if there is a traffic jam that direction can get 6 lanes to speed traffic flow because the direction with less traffic doesn't need all 4 lanes.

 

[thilanliyan]

But then what Josh and Fierydemise said goes against what BenSkywalker said doesn't it?? You can't dedicate say 30 texture pipelines can you (wouldn't that then obliterate all the current gen GPUs??)??

This is what BenSkywalker said:
"No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs."

I'm not sure exactly what he meant there. Anyone care to decode??

 

[ShadowOfMyself]

Originally posted by: thilan29
But then what Josh and Fierydemise said goes against what BenSkywalker said doesn't it?? You can't dedicate say 30 texture pipelines can you (wouldn't that then obliterate all the current gen GPUs??)??

This is what BenSkywalker said:
"No, it can't. It has general purpose shaders for pixel/fragment- vertex and geometry shading based operations. These can not be utilized as TMUs or ROPs."

I'm not sure exactly what he meant there. Anyone care to decode??

No... because they're talking about shaders and not TMUs/ROPs, all he said is the 64 shaders cant act as texture pipes

 

[ElFenix]

Originally posted by: BenSkywalker

anyway, i'm of the opinion that the xb360 runs dx10, or something close enough to not matter, so that ati (which worked closely with MS to develop R500, i'm sure) knows what will be in dx10, so they can get a dx10 R600 out pretty easily

D3D10 offers very, very little in the way of new features. It just does things in a superior way as to not create so many bottlenecks. Don't expect much at all beyond D3D9 in terms of features- it is simply a different approach.

oh i wasn't expecting much in the way of visuals, after all, PS3 doesn't look that much better than PS2 for visuals. my statement goes more to my thinking that ATi has a leg up for dx10 due to their work with MS on the xb360. so they *probably* have a competitive advantage over nvidia as to the next generation

 

[Sonikku]

The Inquirer is teh suxxorz.

Don't believe their lies!

 

[TanisHalfElven]

Originally posted by: Elfear

Originally posted by: bunnyfubbles

Originally posted by: gobucks
to be honest, I'm not sure how much faster R600 will be than the X1900 series. in terms of total shaders, the X1900 has 48 pixel + 8 vertex, or 56 total, while the R600 has 64. That is hardly an earth shattering increase considering the X1900 almost tripled the X1800's shaders, and yet performance wasn't increased by more than 20% in most cases. Also, from what I've read, because of the additional overhead of a unified shader, each one is slightly less efficient than a dedicated pixel or vertex shader. There are advantages, like the ability to dedicate more vertex shaders in very vertex-shading heavy games, but for traditional games, where there is much more pixel work than vertex work, there won't be much of an advantage.

Also, why do we keep increasing pixel shaders without bumping up the number of actual pipes on new cards? We've been stuck at 16 since the 6800/X800 days. What good is having 64 shaders if there are only 16 ROPS to feed them through? I know that a ratio of 3:2 worked well for nvidia with the G70/G71, but there are obviously deminishing returns with anything beyond that (i'm sure the X1900 can't reliably keep even half its pixel pipes busy with its ratio of 3:1), and I'm sure it'll be the same story with the R600's ratio of 4:1

ATI obviously needs your help.

Lol. What were those dense engineers thinking? ATI could save so much money by just employing gobucks as their design engineer/specialist.

All right, sarcasm aside we still don't know the hard specs of R600 or G80, unless I've missed an official announcement. I'm sure there will be a performance increase, how much is yet to be seen.

well if you look at history the top end version (x2800) will be 2xX1900xtx (or what ever R580+ will be called). the lowest end will be equal to today x1600xt.

 

[Polish3d]

Hmm...if they are sticking with 16 texture units they must really believe that's the most optimal allotment .

I read an interview on HardOCP w/ ATI Engineering who said they didnt believe more than 16 TMU would be an optimal use of resources as well as being limited by memory bandwidth IIRC

 

[thilanliyan]

Originally posted by: Frackal
Hmm...if they are sticking with 16 texture units they must really believe that's the most optimal allotment out there.

I read an interview on HardOCP w/ ATI Engineering who said they didnt believe more than 16 TMU would be an optimal use of resources as well as being limited by memory bandwidth IIRC

So even for R600 are there 16 TMUs?? Or does TMUs not apply anymore for the unified architecture??

Anyone have a link that explains this stuff in detail??

 

[hardwareking]

i got a few questions.Doesn't 64 unified shaders mean it's the equivalent of having 64 pixel and vertex shaders?
And how many pipes,rops and tmus will r600 have?And whats the memory bandwidth?Is it still the 512 bit read and 256 bit write?
Does anybody know?