Navigate: Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors WebNews Home	.	.
Prices: CPU Intel P4 2.4C-800 P4 2.6C-800 P4 2.8C-800 P4 3.0-800 P4 3.2-800 AMD AthlonXP XP 1700+ XP 2000+ XP 2400+ XP 2500+ XP 2700+ XP 3000+ XP 3200+ Athlon64 Athlon64 3200+ Athlon64 FX-51 Opteron Opteron 240 Opteron 242 Opteron 244 Opteron 246 Memory Corsair Crucial Kingston Mushkin OCZ Search Prices:

Expendable Timedemo

One benchmark that in the past has mercilessly exposed latencies is Expendable and while we certainly don't care about the absolute numbers, we can still use it as a sanity or double check for what is going on on the memory front. Hyperthreading was enabled in all runs.

As expected, with a single DIMM only, the system takes a substantial performance hit. We tried all different latency settings and there was no difference whatsoever which confirms our SiSoft Sandra results. What is somewhat curious is the fact that with increasing total system memory, the scores increase. Originally we thought that this was an artifact but the effect was consistent throughout the testing.

ContentCreation Winstone 2002

Usually not exactly our favorite benchmark because of the vagaries and the predominant impact of HDD and I/O interface on the score, CCWS2002 turned up some quite interesting results this time. With latency issues out of the way, we can concentrate on physical memory configurations and their impact on performance.

Content Creation Winstone2002 scores using 512 MB of system memory, either as dual 256 MB or else as one single 512 MB DIMM (four bottom results). The two top results use 2 x 512 MB DIMMs. HyperThreading was either enabled (HT+) or Disabled (HT-)

The results are interesting, to say the least. With HT enabled and two 256 MB DIMMs, the performance is abysmally low despite the wide memory interface. Surprisingly, using a single 512 MB DIMM yielded much better performance and did show some minor improvement of HT over no-HT. Going back to two 256 MB DIMMs with HT turned off gave the top score in this configuration. At first, these results are as counterintuitive as they could possibly be, however, there is at least a plausible explanation.

• HyperThreading is seen by the operating system as two independent processors, each of which creates a logical software memory address space which is tagged within the TLBs. That means that two processors are competing for the same address space or, more likely, are dividing the memory address space amongst themselves.
• Moreover, if two DIMMs are used in dual channel configuration, the Granite Bay chipset combines the physical rows of both DIMMs to a single "Überbank" as we described earlier, albeit, at the expense of the memory address space.
• If the two phenomena are combined, that is, a dual channel configuration with half the addresses but twice the width, this address space is further cut in half by the need of allocating it to two logical processors.
• In other words, with HT enabled to simulate two logical processors and two banks fused into one "Überbank", each logical CPU has only 1/4 of the total memory address space available, that is, in a 2 x 256 MB configuration, it amounts to the equivalent of 128 MB system memory address space and that really kills performance. In reality, it is way more complicated but this is as trivial as I can explain it.

As support for our theory, we simply ran the benchmarks again with 2 x 512 MB. If we are right, 2 x 512 MB with HT should give approximately the same score as 2 x 256 MB without HT. Two 512 MB DIMMs without HT should still give the highest score.

QED

There is probably some coincidence that the scores came out perfectly but we won't argue or as they used to say in Latin: quod erat demonstrandum.

next page:    => 3DMark, Comanche4, UT2003 =>