Opinions AMD64 vs EM64T vs Itanium II 2006-08-29 - By Collins, Kevin [MindWorks]
Back
Ken,
thanks for the information. We also like to keep hardware
architectures to a minimal amount, and that ability scale is important
to us. I'll check out the article you've linked as well.
Thanks,
Kevin
-- --Original Message-- --
From: nahant-list-bounces@(protected)
[mailto:nahant-list-bounces@(protected)] On Behalf Of Ken Snider
Sent: Tuesday, August 29, 2006 3:56 PM
To: Red Hat Enterprise Linux 4 (Nahant) Discussion List
Cc: Discussion of Red Hat Enterprise Linux 3 (Taroon)
Subject: Re: Opinions AMD64 vs EM64T vs Itanium II
Collins, Kevin [MindWorks] wrote:
> Just wanted to say thanks for the many opinions that I have so far
> received. In our initial use, these servers will be for serving LDAP,
> VNC, HTTP, etc. Nothing where huge performance is required.
The prevailing wisdom, at least for our applications, is that the
Opteron platform, on AM2, is *far* better suited to Multi-processor (as
you start getting into honking big systems, as in 4-way+). The EM64T
platform will not scale in the same way. Because of this, we decided to
standardize on Opteron, because we didn't want micro-architecture
oddities between 2-way EM64T and 4-way+ AMD64 systems to become an
issue.
There's a *great* Article about this at ArsTechnica here:
http://arstechnica.com/news.ars/post/20060411-6581.html
Snippet here:
"AMD won't be introducing a major new CPU architecture to replace Hammer
until 2007. In the meantime, the company plans to combat Intel's Conroe
by exploiting the edge that their combination of a more scalable,
glueless multiprocessor interconnect-coherent HyperTransport-with an
on-die memory controller gives them over Intel's aging FSB architecture
and off-die memory controller.
For the consumer, these kinds of issues aren't going to have a direct
impact on performance or purchasing decisions. Consumers don't care that
you can gluelessly add new sockets, and that AMD's NUMA design means
that the system's aggregate bandwidth can scale right through the roof.
This is because even high-end consumer-level systems are going to have
one socket, period, and in the single-socket world of content creation
and gaming benchmark bakeoffs Conroe's shared-bus FSB and off-die memory
controller won't hold it back. Not even when the number of cores per
chip increases will the AM2 + HT + on-die DDR2 combination give AMD a
single-socket advantage over Intel, since there's only a certain amount
of bandwidth you can push through one lone socket, HT or no HT.
The main place where the influence of HT + AM2 + on-die DDR2 will be
felt is in the server market. Multisocket AMD boards will be cheaper to
make, and they'll have more aggregate memory bandwidth. Furthermore, in
the realm of four sockets and up, AMD's interconnect and memory
bandwidth edge will probably be decisive."
Basically, we decided to go with an architecture that will scale as we
do.
--Ken.
--
nahant-list mailing list
nahant-list@(protected)
https://www.redhat.com/mailman/listinfo/nahant-list
--
Taroon-list mailing list
Taroon-list@(protected)
https://www.redhat.com/mailman/listinfo/taroon-list
|
|
