[ale] Suggestions for 64 bit dual Opteron motherboards

[Jeffrey B. Layton]

Bjorn Dittmer-Roche wrote:

>On Fri, 23 Apr 2004, Dow Hurst wrote:
>
>  
>
>>I've been looking at performance of Opterons vs Itaniums/Xeons.  The more
>>Opterons on the motherboard the better.  So, 4P and 8P boards are out there
>>now.  Microway or Aspen have 4P units.  A Russian company has created the 8P
>>unit.  The Intel stuff is limited on interprocessor communication by a single
>>pipe while the Opterons have a switch like interconnect with much higher
>>bandwidth.  The Xeon's top out and get starved by the limited pipe they have
>>to talk thru.  It is a long term problem Intel has gotten around for now by
>>bumping up CPU GHz.
>>
>>So you get a real increase in performance on the right kind of code on a 4P+
>>motherboard. ;-)
>>
>>I know your not after that info.  I just wanted to blab about what I've been
>>researching....
>>    
>>
>
>I'm interested! I know this has been a long standing problem for intel.
>One of the big advantages for years of companies like Sun
>and SGI has been their ability to scale processors. Do
>you happen to know how well the Opterons compare to these vendors?
>

Scale in what way? Opterons have a neat memory bus (Hypertransport).
I've only heard of systems going to 8-way, but I've heard rumors of
16-way systems though. The SGIs have a slightly different architecture.
The processors are in a "network" configuration, like a Fat-Tree or
a ring (couldn't think of a better work than "network", but that's not
the right term) instead of having all of the processors on a single MB.
Not sure how they compare, but I personally view them as separate
products (although I'd love to see a 16-way Opteron, but I don't think
AMD has a 1642 ot 1644 - at least not yet).

I tested one of our CFD codes on Opteron (242's I think) and compared
them to Xeons/2.4 (800 MHz bus I think). On our code, the Opterons
were faster by about 90% compared to the Xeons. YMMV.

Enjoy!

Jeff