[Beowulf] Opteron downrates DDR400 to DDR333 performance?

Josip Loncaric josip at lanl.gov
Mon Feb 6 12:30:27 PST 2006

Nate is correct -- but this downclocking of memory also depends on your 
BIOS, which actually determines how to clock RAM.

For example, Opterons Rev. E or later can drive at most 6 ranks of 
memory at full DDR400 speed, says AMD's BIOS and kernel developer 
manual.  This would imply that 4 dual-rank DIMMs would have to be 
downclocked to DDR333.  This is what typical BIOS would do as the default.

However, the actual speed limit depends on electrical loading.  Provided 
that the CPU, motherboard, and RAM cooperate sufficiently well, it isn't 
necessary to downclock memory.  For example, HP's BIOS in xw9300 doesn't.

This issue causes much confusion and I wished that manufacturers could 
explain it better.  Unfortunately, since the actual limit depends on the 
CPU revision, physical motherboard properties, and RAM characteristics; 
most BIOSes follow AMD's conservative recommendation and downclock 8 
ranks of memory even when they could operate reliably at full speed.


Nate Faerber wrote:
> On Thu, 2006-02-02 at 14:05 -0600, NazarioJose at JohnDeere.com wrote:
>> Hello, computer gurus -
>> do any of you know if the AMD Opteron downrates the PC3200 400 MHz DIMMS 
>> (DDR400) to 333 MHz (DDR333) performance  when the system has more that 2 
>> Gbytes of memory installed?
>> i was told that when an opteron socke has more that 2GBytes of memory, then 
>> the cpu will automatically downrate the memory rate to an equivalence of 
>> DDR333 memory.
> It doesn't necessarily depend on the amount of memory installed.  It
> depends on the total number of ranks in your memory DIMMs which is
> likely determined by the number of DIMMs installed.  You get downclocked
> to PC2700 once you hit 8 ranks total on a memory controller (one per
> Opteron CPU.)  So you could get 4GB of RAM on a processor if you use
> single ranked DIMMs.
> See section here:
> http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26094.PDF
> Find previous coverage of this subject here:
> http://beowulf.org/archive/2005-July/013377.html

More information about the Beowulf mailing list