As I understand it, the real problem with centralized forwarding
architectures (and electrical forwarding architectures in general) is memory
access speed. The total throughput of the router cannot exceed the memory
access speed. Even if there was a way to process packets
(lookup/decrement/move across bus, etc) at OC-192 rates in a single
instruction at infinite speed (1EHz, say), the packet must be written into
and out of memory. Given 1ns (possible today?) read/write times, and
assuming 64 byte cell-based packet read/writes for efficiency, you can write
64*8/1ns = 512,000,000,000 bits/sec.
Divide by 2 (read/write) = 256Gbps
Which is equivalent to 25 OC-192s at line rate. Or, 8 OC-768s at line rate.
But since my initial assumptions are currently impossible, then the results
are also not possible using centralized forwarding. For example, by
doubling the memory speed, we lose 12 OC-192s! I would expect even those
vendors that use centralized forwarding engines to go to distributed ones in
order to achieve OC-768. After that, who knows. Memory needs to me
accessed in the picosecond speed range. I think that optical switches will
be in place before core OC-3072 and higher links come online.