Instead of adding dual cores, I would rather
have a larger, more capable cache. For example,
with enough chip real estate, I would rather
have a 16 MB fully-associative L2 cache, rather
than a second CPU. After that, I would look for
ways to speed up CPU memory bus
inter-node communication (e.g. something like
SGI NUMAlink) to get to neighboring nodes.
Someone has decided that adding an extra CPU
is cool, but, I would like to see an objective
engineering trade study comparing the extra CPU
to the option of spending the same real estate
on cache.
Slashdot Mirror
Instead of adding dual cores, I would rather have a larger, more capable cache. For example, with enough chip real estate, I would rather have a 16 MB fully-associative L2 cache, rather than a second CPU. After that, I would look for ways to speed up CPU memory bus inter-node communication (e.g. something like SGI NUMAlink) to get to neighboring nodes. Someone has decided that adding an extra CPU is cool, but, I would like to see an objective engineering trade study comparing the extra CPU to the option of spending the same real estate on cache.