Patrick O'Callaghan <pocallaghan(a)gmail.com> writes:
Actually I'd say that so far it hasn't really panned out, for
the same
reason that general-purpose hardware defeated those special-purpose
machines I mentioned: the market is too small for them so development is
relatively slow, thus the apparent attractions of "doing it in hardware"
aren't enough to compensate, i.e. by the time you get your special
machine out the door the now-standard off-the-shelf parts can knock the
socks off it. The exception that proves the rule is graphics cards,
where there is a huge market that makes the special-purpose hardware
worthwhile.
The other reason is that on-chip caches have gotten so fast that going
to main memory is a huge bottleneck. Even with the caches doing
multi-row reads of the adjacent DRAM memory locations, it still isn't
easy to keep the cache filled for memory intensive operations like
pattern search or block move. Special search hardware in the cpu would
be fixing the part of the problem that is already faster than the
supporting hardware.
-wolfgang
--
Wolfgang S. Rupprecht
http://www.wsrcc.com/wolfgang/ (IPv6-only)