Elmer, <#2 Ali doesn't get it. I pick #2.>
Looks like a wrong pick.
Few other possibilities have apparently
escaped your attention. There are other reasons
why the big server vendors are using Xeons.
First, the scalability of basic PeeII
"multiprocessor" system was limied to
2 (a rather expensive proprietary
non-Intel-made chip set was required
to build bigger configurations) while
Xeons have direct support for 4-way
systems.
Second, Intel promises to increase the
size of L2 cache to 2MB (and up?). This
is a must for a reasonable enterprise
server. The PeeII had no such promise
(I recall some "discussions" about the
L2-size matters for servers on Intel
thread when I said that the PeeII has
ridiculously small L2 cache for servers
while competition made 4 and 8 MB of L2
at that time. Your buddies laughed at
me, but Intel did not, and just did it).
However, Intel has to make their own
SRAM for Xeons, 222 sq.mm, which goes not
for free. In addition, I suspect that the
speed of this so-called "full-speed"
Xeon's L2 is not exactly "full". I think
they fool you by making a 2- or 4-way
interleaved memory to pump up the data
rate, but the latency must be the same
as for regular 1/2-speed SRAM - you
cannot "fool" physics. As Scumbria
has told you here many times, what really
does matter is latency, not the
bandwidth. The better bandwidth sure
helps a bit, but the ROSI on Xeons
is low.
Please note that (inspired by the form
of your reply of who got it and who does
not) I could simply reply that these matters
are well above your fab-technician head,
but I am in the mood today for more
details to educate you.
|
