[Beowulf] CCL:Opteron or Nocona ? (fwd from cavallo at chemistry.unina.it)

[Stuart Midgley]

Morning

It highly depends on your code and whether you are going to run more  
than a single cpu in a box.  If you are only going to purchase single  
cpu boxes, I think they are more or less the same.

If, however, you are going to run more than a single cpu box, I think  
it is fair to say, that the Opterons should give better performance.   
Just look at the spec rate fp base and you will see that a dual  
processor opteron system out perform the Nocona systems by a long way.

Each opteron chip has its own memory controller.  So, if you have 2  
cpu's, you have 2 memory controllers and your memory bandwidth scales  
linearly.  Nocona chips sit on a bus with a single memory  
controller.  If you have 2 cpu's you still only have 1 memory  
controller, so each chip, on average, sees only 1/2 the memory  
bandwidth.

The other issue is memory latency.  Because the Opterons have the  
memory controller on the actual cpu chip, their latency to memory is  
very very low, which gives an effective increase in memory  
bandwidth.  The Nocona chips have the memory controller off chip, so  
their latency is high, reducing effective memory bandwidth.

There are other minor difference, but the memory sub-system is the  
one that gives the biggest performance difference.

Stu.



On 08/05/2005, at 3:23, Eugen Leitl wrote:


> ----- Forwarded message from Luigi Cavallo  
> <cavallo at chemistry.unina.it> -----
>
> From: Luigi Cavallo <cavallo at chemistry.unina.it>
> Date: Sat, 07 May 2005 09:19:33 +0200 (MET DST)
> To: chemistry at ccl.net
> Subject: CCL:Opteron or Nocona ?
>
>
> Hi,
>
> we are going to invest some money for a few computers, and we have to
> make a decision between the AMD-Opterons and the Intel-Nocona. What's
> better ? We are experienced with the Opterons, but we have no idea  
> about
> the Noconas...
>
> Major codes to run on them will be classical QM packages as ADF,  
> G03, TM,
> some AIMD as CPMD, and possibly some classical MD as gromacs.
>
> Thanks,
> Luigi
>
>
>
>


--
Dr Stuart Midgley
sdm900 at gmail.com