Thou shalt have no other gods before the ANSI C standard 1660

David Wagner rate,

Of course you are right in a limited way -- but you should be aware that there is a big shift in the computer industry happening right now. I.e., the shift to parallel processing. If you look at the way that this will happen in an Opteron-like architecture, cores will be clustered into multiple nodes each backed by a seperate memory controller. So the bandwidth basically grows with the number of nodes -- I would expect to see 8-way nodes (with between 16 and 32 cores) fairly commonplace in the next 5 years (beyond that, I would not hazard a guess). There are already Itanium based systems which use *both sides* of the motherboard to place as many nodes into one system as possible. So bandwidth will likely find a way to follow "Moore's Law" (or something like) it.

Of course this will still require quite a lot of compiler work to reduce objects into "stripes" that can be distributed over multiples CPUs, however, there exists proof of concept stuff even today that does this.

Basically that "one CPU can not fill in 64 bits worth of memory-address space in reasonable time" arguement is limited to that fact that far future systems are not likely to have only one CPU in them. I would rather change the argument to one of reasonable data consumption by applications.

--- Paul Hsieh