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:

Test Configuration

*we were using the demo version since we found up to 5% performance variation between individual runs of the retail version benchmarks. The variability appeared to be caused mainly by low response / seek times of the CDROM (having the CD in the drive is a requirement for launching the retail version of UT2003 including the benchmarks). The demo version did not show the same fluctuation in results and even though it is hampered by other limitations it was perfectly suited for the purposes here, that is comparing different memory configurations within the same system only.

Performance

In our review of the Athlon XP2800+, we already showed a ton of benchmarks comparing the different SocketA platforms (VIA KT400) as well as pitting the XP2800+ against different P4 configurations. Therefore, we'll keep the benchmarks short and focus on the different memory configurations and how they influence the overall performance.

We mentioned it earlier, the main crux of the original nForce chipset was the fact that the perforance dropped like a rock if the correct memory configuration was not used. That is, neither a single DIMM nor three DIMMs would even get close to the performance achieved with two single sided DIMMs in slots 1 and 2. We also mentioned that there has been some extensive redesign to eliminate those issues and Sandra is here today to tell us all about it.

SiSoft Sandra Buffering Enabled

SiSoft Sandra 2003 memory bandwidth benchmark scores with buffering and prefetching enabled using 1 or 2 x 256 MB (single sided) or 1 x 512 MB (double sided) DIMMs. Surprisingly, running three 512 MB DIMMs yields the highest memory bandwidth followed by a single 512 MB DIMM while dual DIMM configurations of 256 MB modules or 512 MB in Slots 1 and 3 (not shown) fall behind.

SiSoft Sandra Buffering Disabled

Because of the nature of SiSoft Sandra memory benchmark, differences caused by latencies often don't show up when buffering is enabled. We re-ran the benchmarks with buffering / prefetching turned off. Again, we find that three dual bank modules are reigning superior on the nForce2, meaning that there is definitely no Super-Compatibility or Stability mode kicking in. Keep in mind that the discreets used on all chips were identical (Winbond -5 8 M x 32) meaning that page size or other chip internal features cannot make any difference here.

We can go into a ton of speculations here, citing 24 open pages and whatever other factors as the reason for the performance increase but it is probably much simpler than that. Keep in mind that Sandra will work optimally on "empty" memory addresses which means that as soon as the operating system is loaded a whole lot of data will be already resident in memory. Sandra will work around those blocks which causes less useable memory space per page, more prefetch misses and overall a higher ratio of page misses. Increasing the overall memory space so that the OS load becomes a negligible factor will increase the overall performance. By extension, this is the same phenomenon as the performance degradation that often occurs in repetitive runs of Sandra and which is caused by what is probably easiest described as memory fragmentation.

We did encounter one problem with the A7N8X and that problem occurred whenever we tried to run unbuffered 1GB DIMMs. In that case, the system was still working but performance was cut in half in SiSoft Sandra benchmarks as well as in any gaming application. To be honest, we have to admit that the DIMMs used are built with 32 components whereas the the specs clearly state that DIMMs with more than 16 components are not supported. It should not matter since the DIMMs are stacked such that the chipset should think that only 16 components are present, reality shows otherwise and the fact that performance showed almost exactly a 50% hit does suggest that only half of the chips were accessible causing retries on every other access.

next page:    => ContentCreation WS2002, Gaming =>