Basic Subsystem Tests
Science Mark 2.0 - Memory Bandwidth
ScienceMark 2.0, while not being able to measure bandwidth on both CPUs in a dual Opteron system, doesn't have to in the ZMAXdp, since the memory modules can only be connected to one CPU, CPU0. Dual-channel DDR400 means it should have more available memory bandwidth than the other two systems.
And it does, some 500MB/sec more than the Opteron 250 system and around 72% efficient overall, which is spot on for Opteron.
ScienceMark 2.0 - Access Latency
While the memory modules run at DDR400 in the ZMAXdp, they run at higher latencies than the DDR333 memory in the Opteron 250 system. That said, it should still go out to main memory quicker than the other systems.
Less than 70ns for a trip out to main memory for the CPU's memory controller in the ZMAXdp, quicker by about 20%, matching the difference in memory speeds.
Pifast
Pifast is single-threaded and so gains nothing from having the second CPU present in each test system. The Opteron 250 is acting like a single Athlon FX-53 processor on Socket 940, with the 3.4 Xeon acting like a new 3.4F Pentium 4 on LGA775.
Raw CPU speed helps the Opteron 250 system, some 600MHz faster than the Opteron 244s in the ZMAXdp systems, take the easy win. The ZMAXdp ends up just faster than the Xeons. The performance difference is about right for a combination of faster memory but lower CPU speed.
3DMark 2001SE
Finally, 3DMark 2001SE to round things off in these basic tests. NVIDIA's GeForce 6800 GT does all the hard work, GPU wise, in each system. The Opteron systems use the AGP variant, the Nocona gets a PCI Express version.
The slower CPU speed of the Opteron 244s in the ZMAXdp system are the reason for its difference in performance versus the 250 system, more than any difference in memory subsystem horsepower. We're awfully CPU limited in 3DMark 2001SE with 6800 GT.