You can listen to the video of Urz explaining how Google implement its
system.
It is part of the University of Washington Colloquim series.
It struck me as an example of designing a system that emphasises
intelligence over brute computing power.
The conclusion:
no need for large cache,
no need for byte level parallelism,
only need SIMD, Single Instruction, Multiple Data parallelism(not so
sure about this though)
This lecture is dated to 2002 but is still useful.
The trend is supported by the multiple core cpus. Multithreading was
mentioned as desirable.
