Will R580 have 48 shader 'pipes?'

[Mariner]

Sounds like R580 is going to be a shader-processing monster of a chip!

I wonder how big the chip itself will be? 320m transistors for R520 is a not inconsiderable number so R580 must contain lots of transistors.

 

[ERK]

So after reading some reviews, am I correct to assume R580 will have:

48 vec ALU 1
48 vec ALU 2
48 scalar ALU 1
48 scalar ALU 2

for a total of 192 ALUs of various flavors, bundled into 12 pixel shader processor quads?
Or is that 3 "quad-quads?"

This speculation sure is fun, but the semantics are a major pain.

 

[Rur0ni]

Those numbers make r580 sound like a beast.

edit: You know, had ATI got their parts out when they should have, they'd really be whomping on nVidia right now.

 

[KimB]

Now, what are the die size and transistor counts of the X1300 and X1600 cores so we can get some better idea of how big an R580 core would be?

 

[AlphaWolf]

Now, what are the die size and transistor counts of the X1300 and X1600 cores so we can get some better idea of how big an X1800 core would be?

Er I think you mean r580 core instead of x1800 core (as we already know the r520 die size and transistor count).

If you ninja edit quick I can delete this post.

 

[Jawed]

So after reading some reviews, am I correct to assume R580 will have:

48 vec ALU 1
48 vec ALU 2
48 scalar ALU 1
48 scalar ALU 2

for a total of 192 ALUs of various flavors, bundled into 12 pixel shader processor quads?
Or is that 3 "quad-quads?"

This speculation sure is fun, but the semantics are a major pain.

The ALU counts are right - for shader math. You forgot to count the texture address calculation ALUs, though... An extra 16.

I still haven't the foggiest what Dave is trying to communicate about the shader organisation in RV530 (and hence how it'll scale in R580) so I dunno about the organisation.

Jawed

 

[Colourless]

I would have assumed the 48 Pipes would just be single ALUs, much like Xenos, rather than the current ALU+Mini ALU setup

The transitior increase from 24 ALU+Mini ALU pipes to 48 ALU pipes probably isn't *that* much.

 

[KimB]

One thing to bear in mind is that, as far as I know, ALU 1 and ALU 2 are not symmetric, but rather ALU1 is a MAD unit, while ALU2 is a special function unit (sqrt, and whatever else ATI implements in hardware).

 

[Jawed]

I would have assumed the 48 Pipes would just be single ALUs, much like Xenos, rather than the current ALU+Mini ALU setup

The transitior increase from 24 ALU+Mini ALU pipes to 48 ALU pipes probably isn't *that* much.

We don't know what the detailed ALU arrangement in Xenos is. All we know is that it's vec4+scalar.

The concept of a "mini-ALU" is somewhat strange. NVidia's mini-ALUs are for scaling, clamping and bias. ATI's may do those things, too, but in R420 they definitely do addition.

So counting ALUs is not very meaningful.

Jawed

 

[Dave Baumann]

I would have assumed the 48 Pipes would just be single ALUs, much like Xenos, rather than the current ALU+Mini ALU setup

The transitior increase from 24 ALU+Mini ALU pipes to 48 ALU pipes probably isn't *that* much.

Unlikely the ALU structure will change for two reasons - blocks and wiring for this structure is well known and th compiler is already tuned for it.

A for what is occuring, all I would say is that at this part of the pipeline they are tripling the number of quads it operates at.

 

[Dave Baumann]

ALU breakdon is in the article:

http://www.beyond3d.com/reviews/ati/r520/index.php?p=03

 

[PeterAce]

So each Quad PS Core will gain a second and third set of the current ALU structure (ALU1, ALU2). Meaning that R580 will still have 4 Quad PS cores (but each core will have more ALUs packed in).

 

[hkultala]

Now, what are the die size and transistor counts of the X1300 and X1600 cores so we can get some better idea of how big an R580 core would be?

about 100 to about 150 M.
and they also increased vertex shaders from 2 to five, and Z-units from 4 to 8.

by guess is about 440Mtransistors for R580.

 

[Dave Baumann]

RV515 is a little over 100M transitors, RV530 is 157M. Bear in mind that RV530 also has a far more complex memory controller/interface and double pumped Z/Stencil in relation to RV515 (and more VS and Chalnoth points out) - given the size of he memory controller in R520, the number of transistors in RV530 dedicated to this shouldn't be underestimated.

 

[Dave Baumann]

So each Quad PS Core will gain a second and third set of the current ALU structure (ALU1, ALU2). Meaning that R580 will still have 4 Quad PS cores (but each core will have more ALUs packed in).

Yes and no.

Its a little complex how I view it, but just saying "4 quads with more ALU's" implies that the ALU's are deep (i.e. lots of instructions executed on a single pixel), but they not, they are wide (lots of pixels). However, conceptually I still see ATI's Quad pipes as a single quad pipe - from intput, to shader, to ROP - in the case of RV530, though, is pushing three quads of pixels through the ALU area at any one point in time (you could say "Triple data rate" through the ALU's).

 

[PeterAce]

Yes and no.

Its a little complex how I view it, but just saying "4 quads with more ALU's" implies that the ALU's are deep (i.e. lots of instructions executed on a single pixel), but they not, they are wide (lots of pixels). However, conceptually I still see ATI's Quad pipes as a single quad pipe - from intput, to shader, to ROP - in the case of RV530, though, is pushing three quads of pixels through the ALU area at any one point in time (you could say "Triple data rate" through the ALU's).

Thanks Dave, thats much more clear now.

 

[Jawed]

So one batch in RV530 is 48 fragments, processed in four phases, 12 fragments at a time?

Meanwhile 4 fragments can be textured at the same time.

While the 48-fragment batch is waiting for texture results, other 48-fragment batches can start processing in the shader (each of those proceeding until texturing is required).

So in R580 are we going to see RV530 scaled-up 4-fold? Will each batch consist of 192 fragments, running in four phases, through a 48-pipe shader? With the texture pipes capable of operating on 16 fragments?

If R580 doesn't use phased batch processing, that would make each batch 48 fragments. Which seems much more reasonable.

Jawed

 

[Dave Baumann]

Added some clarifications and diagrams to better explain the pipeline differences between R520/RV515 and RV530, here.

 

[Megadrive1988]

R580 will have either 16 or 24 pixel pipelines (and ROPs) , but within the 16 (or 24) pipes, there is said to be 3 fragment shader ALUs per pipe, instead of 1.

if R580 has 16 pixel pipes, then it'll have 48 fragment shader ALUs

 

[no-X]

Die of RV530 is 43% larger, compared to RV515 (xbit-labs pictures). Could it be similar for R580/R520?