Ok, Your not understanding the concept at all. I'll do my best to talk threw what your missing. I'm not going to repond to duffers extreme pestimism quote-by-quote as thats not helping. So, just hang with me.
Basic idea. You get a whole bunch of existing or near future technolgies and put them together. Hopefully, the sum will be much greater than the parts.
Overview (go into more details later) You get a fast and scalable processing architecture -
CELL - that runs off a single OS, conforms to OGSA GRID computing standards, and has other cool features like many autonomous aspects (self-enhansing preformance, securitry, ect).
The network is thus boadband based and can be used for gaming, distributed media content - basically anything that used offline processing power.
But, this is Beyond3D, so I'll concentrate on games. The idea is to compute all raster functions locally (where they must be), but do what Sony's pushing as "
World Simulation" on the network side. This can be anything that doesn't require RT processing like rasterization - Think AI, Physics, basically once you get used to the idea you can go a hell of alot.
So, by splitting up the tasks you free, say, ~50% of the local resources; but double the local processing potential devoted to rasterization and increase the 'World Simulation' stuff by a theoretically infinate level. This is how you beat Moore's Law by 10 years, you need to boost transistor count and do it threw GRID computing on offlines tasks.
The costs are absorbed by (just me thinking) the distributed content - Sony realigned itrself in 2000 and is preparing for convergence and the digital distrobution od media from their huge production houses. (They have like 10 or 100X more digitized film and TV than the next person). But this really doesn't concern Beyond3D.
Now, He expressed concern of Cell's architecture and bandwith of 32 CPU's sucking threw one bus. Well, According to Okomoto at GDC '02, the Cell is composed of the Cores and a "pluralty" of RAM. The whole idea of Cellular computing is keeping the execution units fed by integrating the whole "cell" or execution units and their assocaited RAM and busses. So, on-chip bandwith should be huge.
The idea behind Cell as off 2001 was to totally kill and end the idea of a cache and put everything on the die. I don't yet know the full microarchitecture oc "Cell' and can't say how they intend to get 1TFlop as Kutaragi said. They licenced the MIPS-64 core, but unless they strip it down and bolt on some serious FMACs I dunno... yet
They stated 1 TFlop, so we'll see.
I think thats enough information for now
It's about time they get newer graphs
PS. If anyone who I sent the old ones to has webspace and wants to post them feel free to - because I don't have space... just label it Vince's present to the asshole naysayers and I'll be happy.
