R520 die size and transistor count redux

[RoOoBo]

80-100 M as each quad block seems to be roughly something like 1/7 to 1/8 of the die area.

 

[Jawed]

Yay, nice!

As a matter of interest, I can find bits of E surrounding the "lone" C towards bottom right. Just not all the bits... But I suspect there is a fourth E on the die, just spread around as Farhan originally noted.

Jawed

 

[Farhan]

Yes, ABCDE seem to come together, although i wonder why they did not place the 4th one at the bottom right corner to make it all nice and symmetrical heh.

 

[Maintank]

The original slew of "R520 = 32pipes" rumors have their origin in R520 die size. It has been know for some time that it would be rather large ( > NV40), leading some to conclude that it will indeed be 32-pipes. As it stands, R520 is the biggest graphic chip to date - at 320Mts, its twice the size of its predecessor. So, what are some of the key factors contributing to this massive transistor count increase? Memory controller? SM3 logic? Larger Caches?

I remember seeing somehwere ATI mentioning moving from FP24 to FP32 would require 50% more transistors. You can then tack on SM3, enhanced memory controller, cache enchancements, and other things you can probably see how it balooned to twice the transistor count as the R420.

 

[rwolf]

Take a look at page 9.

Some may question why the X1800 XL's are being introduced immediately, but the XT's coming a month later. Silicon for all the parts are in production, but the final configuration of the silicon and metal layers was resolved fairly late, and with the production times taking up to three months for chip orders to final products the XT's are coming a little later, once those chips appear. ATI can make XL's available fairly shortly because they had placed large silicon orders earlier, but stopped the production once they realised there were still issues, meaning there were many cut wafers but without metal layers – as the issues could be resolved with a change to the contacts and metal layers ATI could utilise the silicon that had already been cut and apply new metal layers and all of these are going towards the initial XL products.

Does this mean the sample used for previews is not the FINAL X1800XT?

 

[Hanners]

Does this mean the sample used for previews is not the FINAL X1800XT?

As I understand they are final boards, but ATI are holding back a retail release until they have the boards available to them (and AIBs) in large quantities. In other words, they're basically waiting while they build up their inventory.

 

[Jawed]

http://www.beyond3d.com/forum/showpost.php?p=587539&postcount=94

There appears to be a "mix" of XTs out there...

Jawed

 

[Tim]

Those marked areas are 28% of the die, by the way - roughly.

That would put the R580 at 500 milion transistors, I think that is a bit to the high side. I would really like a die shot and transistor counts for the RV530 that would put us in a much better position to figure out what is exactly is what on that core.

 

[Sunrise]

Current XLs seem to be marked as "A14FG", while XTs are marked as "A15FG".
Still something doesn´t make sense:

Why does the XT consume that much power compared to the XL. It could be because of 8 chips with "-BJ12" which indicates 2.0V-operation, but that much more ?

http://www.xbitlabs.com/articles/video/display/radeon-x1000_14.html

112 Watt vs. 59 Watt ? A 100% increase in wattage ?
So this is a pre-overclocked (metal fixed) XL-core ?

 

[hkultala]

There is also 2 times same block, mirrored vertically on right of the downest A/D blocks.

 

[_xxx_]

I'm willing to bet that that center block is cache memory for local register storage.

No, it's embedded DRAM!

 

[_xxx_]

EDIT: deleted, need some sleep...

 

[Tim]

112 Watt vs. 59 Watt ? A 100% increase in wattage ?
So this is a pre-overclocked (metal fixed) XL-core ?

The core should use (2^2)/(1.8^2)*625/500 = 1.54 times more, with twice the RAM for XT the difference is not completely out of line. Anyway 2.0V (and even 1.8V) seems a bit extreme for a 90nm chip - is it the TMSC process that requires high voltages or is Ati still having trouples reaching the needed clockspeed (I would expect something like 1.4-1.6V for a 90nm chip).

 

[Jawed]

That would put the R580 at 500 milion transistors, I think that is a bit to the high side. I would really like a die shot and transistor counts for the RV530 that would put us in a much better position to figure out what is exactly is what on that core.

Depends on what those red, green, blue and magenta areas are, though? I'm guessing that they are pixel shaders and texture pipes.

R580 will only have triple the pixel shaders, not triple the texture pipes. So, which bits are the texture pipes (are they part of the RGBM areas?)?... Where are the vertex shaders?

Jawed

 

[Tim]

Depends on what those red, green, blue and magenta areas are, though? I'm guessing that they are pixel shaders and texture pipes.

R580 will only have triple the pixel shaders, not triple the texture pipes. So, which bits are the texture pipes (are they part of the RGBM areas?)?...

Yes you are most likely right - that would put us around 440 million for the R580, I have said max 450 milion for the 90nm process before so I can accept 440million. My current guess (and I am sure that I am wrong):

A = pixel shader quads
D = Texture units
B = Ring stops
E = Register arrays
C = Threading engines
(feel free to call me an idiot)

Where are the vertex shaders?

Beats me, I guess they have to be somewhere i lower right corner.

 

[vb]

- B should be ROPs: I think the big center block should be able contain all ring bus bits.
- texture units are decoupled, so there is no reason to have them placed near pixel units.
- and where is the already famous "Ultra Threaded Dispatch Processor"?

V

 

[Tim]

- B should be ROPs: I think the big center block should be able contain all ring bus bits.

Yes that sounds right.

- texture units are decoupled, so there is no reason to have them placed near pixel units.

Actually Dave said that it still was 4 texture units for each shader quad, not 16 shared between all shader units.

- and where is the already famous "Ultra Threaded Dispatch Processor"?

It is not really one big threading engine, but one small for each quad.

 

[Jawed]

I agree that "texture units are decoupled" in the sense of batch scheduling - but it seems that each shader quad "owns" a texture quad.

A die shot of RV530 would be very much appreciated, as there is no ownership of texture quads by shader quads...

Jawed

 

[Geo]

Yeah, that memory controller on 130nm wouldn't have left much for it to control, would it?

I look at that and think ATI does have their NV30. . .from an architectural shift pov that will continue to impact their chips for years to come. Only theirs is at least competitive, and sometimes superior, to the competition immediately --it's a better baseline, so to speak.

 

[Kaotik]

Current XLs seem to be marked as "A14FG", while XTs are marked as "A15FG".
Still something doesn´t make sense:

Why does the XT consume that much power compared to the XL. It could be because of 8 chips with "-BJ12" which indicates 2.0V-operation, but that much more ?

http://www.xbitlabs.com/articles/video/display/radeon-x1000_14.html

112 Watt vs. 59 Watt ? A 100% increase in wattage ?
So this is a pre-overclocked (metal fixed) XL-core ?

Could it be, that the current XT's draw so much power to somehow work around the "sucky" core revision? The way I understood it, all the current cores are the ones that got stalled and "fixed on the fly" or something