[quote]Originally posted by Rasputin:
Well, I sort of extrapolated the system for calculating bandwidth from graphics cards articles I've seen. That is, the hertz rating times sixteen. I'm sure you know it better than I
It must be gigabits
per second, I suppose.</strong><hr></blockquote>
is correct... here's why:
The x16 is because many graphics boards use a 128-bit bus. Some of them may even be using 256-bit buses... although the memory and bus usually have a fairly radical organization, which is possible due to the nature of graphics (which by its nature is very highly parallel and very deeply pipelined). The memory is also tightly coupled to the graphics chip & not expandable. Think of it as 4 x 64-bit busses, rather than one 256-bit bus. The latest nVidia highest end graphics cards manage a little over 10 GBytes/sec (as you mentioned) by using this mechanism.
This kind of configuration doesn't really work for the main CPU, although with the advent of onchip memory controllers and per-CPU memory some of this will move to the motherboard. If Dorsal's description of the system is accurate then a dual G5 machine, could
be built each with its own 128-bit bus to its own private memory, would effectively have a 2 x 128-bit memory bus. If the memory is kept very close to the processor(s) then it might be possible that the bus can be wider and/or faster than 128-bit 133 MHz double pumped, but I won't speculate on that.
Adding more processors effectively widens the bus from a system viewpoint, although not from a single processor's viewpoint. This organization has a significant impact on the operating system, however, and it will be interesting to see how MacOSX deals with that. Processor's accessing eachother's memory would see a severe speed penalty -- on both processors. This would mean that either processes need to be bound to processors (limiting), and/or a sophisticated paging method for transfering whole virtual memory pages between processors is required... doing it at the cache level would involve a great deal of overhead.
It Dorsal is accurate and Apple gets the OS right, these machines will be very fast
-- especially in multiprocessor configurations.
I should also add that they could also build these machines with two processors on the same daughtercard, sharing the same memory. This is probably more likely, but without knowing more about the on-chip memory controller its hard to say how it would work.
[ 04-07-2002: Message edited by: Programmer ]</p>