R580 will be called as Radeon X1900

[rwolf]

R300 and up can do 3 ops per pixel shader max. Full+Add+Vector

R580 will do 3 ops per pixel shader pipe x 16 pipes x 3 pixel shaders per pipe = 144 ops max.

G70 has two Full ALUs per shader pipe, but I understand one is shared with the TMU. R520/R580 have separate TMU with separate processing.

[tertsi]

What about GTX with 24 pipes & 2 full ALUs per PS unit ? So, we've 48 ALUs here. If nvidia mods the G71 for 32 pipes, then 64 ALUs. Still far away from the 96 ALUs in the R580, if i'm not wrong. So, for large programs, the R580 should be more efficient. But for short programs, the G71 should be faster due to all the PS units will be busy.


G70 has also special function unit (SFU - NRM, RCP, RSQ etc. instructions) per ALU (shader unit1 and shader unit2). So, total is 2x MADD ALU + 2x mini-ALU + 2x SFU per fragment (or pixel) pipeline. Build-in texture processor (TMU) in first ALU (shader unit1) can work independent (e.g. texture fetch doesn't affect to shader performance).

16-1-3-1 could be also mean 3 x ALU cores per pixel shader pipelines (one ALU core has 2 vector ALU's + 2 scalar ALU's). Still we don't known ALU's types - will it be full MADD ALU and ADD ALU or two MADD ALU's? - I'm thinking it will be full MADD ALU and ADD ALU because two full MADD ALU's eats lots of transistors :)

Now lets put rumored specs again on the diagram:

Possibility A)
R580 = 4 quad pixel shader cores (total 4) = 4 pixel shader pipelines per shader core (total 16) = 1 ALU core per pixel shader pipelines = 3 vector ALU's (+3 scalar ALU's) per ALU core / pixel shader pipeline (total 48 vector ALU's (+48 scalar ALU's))

OR....

Possibility B)
R580 = 4 quad pixel shader cores (total 4) = 4 pixel shader pipelines per shader core (total 16) = 3 ALU cores per pixel shader pipelines = 3 vector ALU's (+3 scalar ALU's) per ALU core / pixel shader pipeline (total 96 vector ALU's (+96 scalar ALU's))