We were in a conversation with a reader who runs a large dual Xeon cluster and he pointed out that data intensive programs scale well on Opterons, but not at all well on Xeon architecture. In fact, there are several threads on the BEOWULF.ORG site which discuss this matter.
Now we've had an interesting letter from a reader who has taken the time to explore how true this contention is.
He has compared single to dual Opteron systems by looking at these SPEC results 14.1 SpecFP 2000 and 26.7 SpecFP 2000 here, which he says shows a a 1.89 speedup, on what he describes as a fairly memory bandwidth intensive benchmark.
This reader uses an earthquake simulation app and says that shows the following results:
Single CPU+Hyperthreading P4c-2.6 800 Mhz FSB:
One job at a time 82.20 seconds
Two jobs at a time 138.21 seconds
Scaled 1.18 times.
Dual p4-3.0/533 Mhz FSB (fastest dual p4 made)
One job at a time 84.98 seconds
Two jobs at a time 155.92 seconds
Scaled 1.09 times.
Dual Opteron 240s (the slowest Opteron)
One job at a time 97.09 seconds
Two jobs at a time 98.77
Scaled 1.96 times.
He adds: "Basically the opteron scales amazingly well. I was shocked to find a 2.6 GHz Single CPU HT to be faster then a dual CPU P4 3.0 Ghz."
He concludes: "The slowest Dual Opteron is 1.57 times faster (wall clock time) then the fastest dual Pentium 4". µ
Sign up for INQbot – a weekly roundup of the best from the INQ