I'm not expecting Ms taking this route but if SPUs shine at something it's perfs per Watts they are pretty outstanding in this regard. I would bet they grossly outdo GPU at "mundane" calculations.
I don't expect either SOny to take that road but I've a feeling that something really "outlaw ish" in regard to nowadays standard hardware designs could surprise, in efficiency and performances.
(I think of those posts from
Fafalada and
Panajev2001a as back up for the assumption).
mmm.... indeed. I do wonder how the current gen CPUs actually compare to more recent designs in the low power sector (Atom, Llano, ARM). Of course, even 45nm XCPU or Cell are still going to be an order of magnitude greater in power, but what if one scaled those particular designs up
Another things so we look @ late 2013 or 2014 for next generation hardwares, it kind of sucks...
(or better be impressive).
At least we'll see what happens on the memory side of things. It's been pretty boring with GDDR5 for awhile. No hints on the next iteration.
At least as far as the 360 design philosophy went, it would be rather interesting to see the CPU be attached to the GPU right off the bat (because it'd start from a low-power design), and they could budget more for a potential off-die eDRAM.
-----------------
On a side note, regarding the job listing for future 360 developments (
3rd link), I suppose we're not really done with the eDRAM. Based on measurements, we're still at 65nm for the Slim.
http://forum.beyond3d.com/showpost.php?p=1447163&postcount=239
The expected scaling from 64mm^2 @ 80nm to 65nm would be 42mm^2. We got ~45mm^2. Fair enough. Nothing is perfect, plus there are the ROPs and such to build. A hypothetical shift to 40nm would put that down to around 17-20mm^2. (
TSMC 40nm production began in 2010)
hm.... A bit curious the article mentions 45nm, not 40nm for the CGPU. Anyways, scaling up for next gen using a similar design would still be... peculiar to imagine.
a) double or triple the ROPs, increased functionality e.g. texturing
b) quadrupling die size (@ imaginary 40nm) would bring it back to 360's original 80mm^2 for eDRAM @ 90nm.
30-40MB eDRAM @ 40nm, ~80mm^2
Give it a few more years for 28nm to ramp up... Could maybe hit 70-75MB @ 80mm^2 *shrug* Fun to think about it.
-----------------------------
Self-reminder to read sometime