MikeB said:
Less than 40% in theory regarding max potential, in reality the Xenon is less efficient as it has to share L2 CPU cache between all three cores, each Cell SPE has super fast dedicated RAM to its disposal, making the design near 100% efficient (other common CPU architectures become increasingly more inefficient by adding more cores). Of course within the 360 design regarding the main RAM the Xenon also has to share bandwidth wth the GPU and using its main RAM it suffers far more from latency. The 360 CPU is far more like a normal desktop PC CPU than the Cell is, apart from also being an out of order processor like the Cell and its cores being nearly identical to the Cell's PPE the general design and layout is very different compared to the Cell. It would yield no performance advantage using it compared to a fast PC processor (only disadvantages). |
I'm sorry, but this is just wrong.
"Less than 40% in theory regarding max potential, in reality the Xenon is less efficient as it has to share L2 CPU cache between all three cores"
The task is not memory bound. It is FP bound. The Cell's much higher FP performance accounts for the vast majority of the advantage. I don't know where you got that 40% from - are you just counting number of cores? Because the SPUs are wholly different and can't be counted like that.
"each Cell SPE has super fast dedicated RAM to its disposal, making the design near 100% efficient (other common CPU architectures become increasingly more inefficient by adding more cores)"
The current generation of PC CPUs isn't bottlenecked by I/O and scale close to 100% when you add cores. Cell isn't more efficient any more - it lost that advantage when Intel added an IMC with Nehalem (AMD had one since Athlon 64).
"Of course within the 360 design regarding the main RAM the Xenon also has to share bandwidth wth the GPU and using its main RAM it suffers far more from latency."
XDR memory may have a higher max bandwidth but MUCH higher latency than the GDDR3 used in the 360. I imagine for this kind of non-memory-bound task that the two factors balance out. Sharing with the GPU is no problem in this case because it is only the CPU that is being stressed according to the article.
"The 360 CPU is far more like a normal desktop PC CPU than the Cell is"
Yes... mostly in ease of programming and flexibility.
"apart from also being an out of order processor like the Cell"
It's in-order. For a single-purpose program that probably gives it an advantage over a desktop CPU as if the instructions are properly reordered and optimised in the compiler for that CPU (as they should be) then being in-order saves on pipeline length which can increase execution speed (e.g. by reducing the penalty for incorrect branch prediction).
