I wrote:
> > Come to think of it, factoring would be an excellent
> > application for a 21264 without an L2 cache, but I
> > don't know if they come that way (except perhaps in
> > a massively parallel setting, where the problem of
> > maintaining cache coherency in a multilevel cache
> > hierarchy often is nearly intractable).
Brian Beesley wrote:
>Either the L2 cache is on the chip, a la Celeron/PIIIE,
>or it's on a board - either a cartridge slot, a la PII,
> or on the mainboard, a la Pentium Classic. If the L2
> cache is off chip, then clearly systems not containing
> L2 cache can be manufactured. If it's on chip,
> presumably you can still turn it off - though you're
> going to be stuck with the manufacturing cost.
For the large and variable L2 cache sizes of 21264
systems, the only reasonable way to to it is off-chip.
If you look inside a 21264, you'll see the daughtercard
with the CPU+monster heatsink, and the separate bank of
L2 cache chips, with their own heatsink covering the
whole batch (but I think only the CPU has a cooling fan-
for the cache, the heatsink plus inside-case air
circulation appears adequate.
Assuming a nominal 1 ns CPU cycle time (i.e. 1 GHz CPU),
the fact that the speed of electrical impulses in copper
wiring is ~2/3 the speed of light in vacuum, or 2x10^8
m/s, if the furthest part of the L2 cache is separated
from the CPU by (say) 20cm of wire, the absolute minimum
number of cycles to fetch data from the cache is
precisely one. If one's CPU architecture is designed to
go to (say) 10GHz, and one is willing to wait up to 10
cycles for data from the L2 cache, fine- beyond that one
must either move the cache on-chip or in some other way
much closer to the CPU.
Within the next 10 years, I fully expect we'll start
seeing compact 3-D stackable-layer designs: processor(s)
plus layers of high-speed memory, with a dense network
of microchannels to carry coolant. Not so different,
really, than what has evolved in the higher animals.
(One common theme of mechanical design and natural
evolution being: If it works, use it!)
-Ernst
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
