silent_guy said:
Hmmm. There's definitely a possibility that there are a number of excess units for redundancy. But an overhead of 33% (because that's how you should look at it!) is out of the question.
[Sadly the 1337 mod skillz of one of our moderators, Vysez, meant that my follow-up post on this subject got moved out of context:
http://www.beyond3d.com/forum/showpost.php?p=712829&postcount=7 ]
I'd be interested in your interpretation of the following die shot (ATI's Xenos for XBox 360):
I have coloured in what I think are the four shader arrays of this chip - each array appears to be split in two, either side of what I guess is scheduling/queueing/back-end type functionality.
Each shader array consists of 16 pipelines (rather than the conventional 4 pipelines that are normally considered to be a "unit" within a GPU). The texturing functionality of these pipelines is entirely separate as a single 16-pipeline unit - somewhere (I guess it's that big pseudo-symmetric block at the right of the die, just above MC0).
The chip is supposed to be around 232M transistors, and each shader array appears to consume about 8% of the die.
Finally, this GPU offers only 48 active pipelines - yet the design appears to consist of 4 arrays, each of 16. i.e. though architected for 64 pipelines, 16 appear to be dropped for redundancy.
In summary, only 8% of this die is consumed by shader array redundancy - if my theory is correct.
---
So I'd be interested in your interpretation of the layout/redundancy of this die, which appears to use 33% redundancy for the pipelines - in total the four shader arrays consume about 33% of the entire die.
---
As a matter of interest, at a guess, R580's fragment shader pipelines (if there are really 64, with 16 dropped for redundancy) would consume about 33% of the entire die: 128M transistors out of 384M.
Jawed