Michael S (already5chosen@yahoo.com) on 11/10/08 wrote:
---------------------------
>EduardoS (no@spam.com) on 11/10/08 wrote:
>---------------------------
>>Henrik S (henrik.stahl@nospam.oracle.nothanks.com) on 11/10/08 wrote:
>>---------------------------
>>>Interesting enough, the NEC submission has a higher peak score (288502 vs 267804)
>>>despite the fact that it is uses a smaller Java heap and the slower (?) 4 GB DIMMs.
>>>This is counterintiutive to me and makes me suspect that the AMD Xeon submission
>>>is on a machine with a less optimized BIOS.
>>
>>Registered DIMMs have a significantly lower latency than FB-DIMMs,
>
>True

And lower power as well.

>
>>also FB-DIMM's
>>latency increase with more DIMMs per channel.
>>
>
>Not in this particular case (8 DIMMs=2 DIMMs/channel). The effect starts from 3-4
>DIMMs per channel. 2 DIMMs per channel were considered optimal on Blackford. Don't
>know if the same is true for Seaburg but I don't see why not.
>

Flat wrong. FB-DIMM latency goes up with each FB-DIMM in a channel as the data passes from one FB-DIMM to the next sequentially. So you go from a one hop latency to 1.5 hop latency on avaerage, from 1 to 2 FB-DIMMs.

Also note that for 16GB configurations, all SPECPOWER Xeon systems have lower results than the Opterons of equal speed. The 2384 (2.7GHz) has the highest result at that amount of memory, 860 versus 843 (2.8GHz Xeon quad).