the following is all quote from the message board thread: 
http://www.ga-forum.com/showthread.php?t=51518

quote:

 "This is actually part 3 of the Watch Impress Kutaragi interview."
http://pc.watch.impress.co.jp/docs/2005/0613/kaigai189.htm

 "He talks about why a Cell-based GPU wasn't use, the close bidirectional 
Cell<->RSX relationship, eDram, unified shaders vs dedicated shaders, why 
they went with NVidia, backwards compatability and more. Unfortunately I 
can't see any proper translation yet, so we'll have to make do with the mess 
that is babelfish's for now."

Things I took from it:

- PS3 backwards compatibility is through software on Cell (perhaps Cell 
alone). Cell has been made bi-endian to emulate the PSX and PS2 CPUs.

- A 2-Cell PS3 was considered - one acting as a GPU, one as a CPU. But 
Cell's nature as a CPU eats away at its efficiency for graphics relative to 
a dedicated GPU. So they've married Cell with a specialised GPU.

- Cell, however, can be used for graphics tasks. Cell and RSX have direct 
access to the results of each other's operations. Kutaragi specifically 
mentions displacement mapping as a sample workload the SPEs could take on. I 
think he mentions something about harmonising floating point rounding errors 
between the CPU and GPU to further ease data exchange between the two. Seems 
like they focussed quite a lot on harmonising communication between the GPU 
and CPU, and binding them quite tightly together.

- Originally there was no VRAM in PS3. The graphics chip accessed only the 
XDR ram.

- There is no eDram in PS3, because they couldn't put enough on the die to 
support a 1920x1080 frame let alone 2. Kutaragi thinks the transistor cost 
of eDram is too high when you consider what the same amount of transistors 
buys you elsewhere in terms of more shading power.

- He touches on unified shading and mentions something about problems with 
stalls..I don't know, hard to make out exactly what he's saying there.




Quote:
      The reason which does not adopt the Cell based graphic chip

      With PS3 architecture, those where last year you are surprised 
graphics are to apply with Cell based architecture. Why, not making Cell 
based GPU?

      Kutaragi: 7 SPE of Cell (Synergistic Processor Element) it can use in 
graphics. Actually, several of the demonstration with E3, when it has not 
been the graphic processor, has done everything of graphics to rendering 
with just Cell. But, such using is wasted. There are times when you do more 
in other things in Cell.

      2 placing Cell, (you use Cell of one side) with also the plan which is 
said it was on the graphic center, but as Cell and Shader as the computer 
thinking that weight is different from function you stopped. We would like 
to make the architecture which Shader with, all things thoroughly (in 
graphics) can specialized Shader. Though, for example, displacing mapping 
(it does) with it can also say with SPE.

      So far (real time 3D graphics), that it seems and shows, (3D graphics) 
in the space is different really. Even then, with present resolution it was 
good. The game which has come out with Xbox 360 the majority, is 3D such as 
that even at present point in time.

      But, I would like to make 3D where (metamorphosis and the like is 
reflected securely in the 3D space). For that, there is the thought as much 
as possible of liking to share the data (with CPU and GPU). Therefore, the 
latest architecture was taken. Originally if as for the floating point unit 
of GPU and Cell, from precision round (making round) until error we would 
like to make entirely the same. This time, it has become close, almost 
rather simultaneous. Therefore, each other, it can use (the operational 
result) bidirectionally.

      - EDRAM is stopped because of full HDTV

      EDRAM (it installed graphic memory and DRAM) expected, but the HDTV 2 
picture hearing, you could agree upon the reason which it does not make 
eDRAM.

      Kutaragi: Originally, there is no graphic memory in GPU, Redwood (the 
high-speed interface which connects Cell and RSX) with YDRAM (the code name 
of XDR DRAM) with also correspondence is achieved. Because YDRAM (memory) 
has become unified.

      But, so entering, that there is a problem which you say whether () 
definite form processing and Shader of graphics (computing) when it does, 
the distant place (accessing memory) with, it is possible to make (zone and 
cycle time) wasteful. It is not necessary to use (memory zone), with 
definite form task of the side of with special care Cell. Because Shader 
with Shader, calculates the tremendous quantity, there is a necessity of 
memory even to there. Especially, when with full HDTV, 2k