Let's start with memory performance tests in RightMark Memory Analyzer, a test package developed by our programmers. You may remember that the engineering sample of a motherboard on nForce4 SLI IE was strongly outperformed by i925XE in real and maximum memory read rates. NVIDIA fans may relax now: the production-line motherboard did not confirm this tendency and it even slightly outperformed the i925XE representative. Note that nForce4 demonstrates practically the same performance with 2xDDR2-533 and 2xDDR2-667. It matches the performance of i925XE+2xDDR2-533 and its maximum value is actually limited by the FSB throughput (in fact, it's even a tad higher (?)). In case of single-channel mode, the read rate of nForce4 SLI IE is predictably limited by the memory rate (FSB throughput is higher), you can evaluate the lag from the dual-channel mode on your own. Note that single-channel modes are not very interesting from a practical point of view, but as our article is devoted mostly to testing the theory, we have carried out a full cycle of tests in these modes for the NVIDIA's chipset.

The maximum write rate in all the chipsets [in dual-channel mode] is limited by the CPU architecture; there is nothing interesting here (the limiting factor of the single-channel mode is again the memory module bandwidth). But the real write rate of nForce4 SLI IE turned out slightly higher than in Intel chipsets, so NVIDIA can celebrate victory (the reference board was on a par with i925XE): its performance with DDR2-667 is higher by ~23%, with DDR2-533 — by ~17%.

Last time we had a draw in memory [pseudo random] read latency, but now the NVIDIA chipset shoots forward (approximately by 2—6%), it outperforms i925XE not only in dual-channel modes, but also with single-channel DDR2-667. It's only slightly outperformed with single-channel DDR2-533.

The interesting effect demonstrated during previous tests, when nForce4 SLI IE outperformed the Intel chipset in random memory read latency (being on a par with it in pseudo-random memory read latency), can be now seen only in a single-channel mode with DDR2-533.

Let's draw a preliminary conclusion based on low-level memory tests: NVIDIA nForce4 SLI IE is actually faster than i925XE; DDR2-667 outperforms DDR2-533 a little at the same low timings; we are inclined to take i945P results as preliminary, so we are not going to rely on them in future (what really kills our interest is increased timings for 667MHz memory, which nullify potential advantages of the increased bandwidth). Now, let's have a look at real applications and evaluate the difference between system results for regular users.

The difference between Intel and NVIDIA chipsets in archivers is totally contemptible (about one percent to the advantage of the latter) and is quite comparable with the natural spread in test results. Trying to take this difference seriously, the i925XE goes on a par with nForce4+1xDDR2-533, while a motherboard on NVIDIA's chipset with single-channel DDR2-667 is insignificantly faster.

As archivers didn't show any differences, we can safely skip tests without 3D graphics, because the other applications that we use are less dependent on memory performance. Indeed, as you can see in this case, video encoding results are the same for all contenders. And compiling a PDF file in Adobe Distiller provides a significant chaotic spread in results (due to continuous access to HDD).

SPECviewperf (a benchmark for professional OpenGL application performance) is very useful because its tests react differently to the performance of system components, so that you can almost always notice interesting patterns. Today, three tests turn out critical to the memory bandwidth, which leads to the victory of nForce4 SLI IE with dual-channel DDR2-667. In other cases the chipsets are either on a par (3 tests), or outperform Intel chipsets (including i945P, 2 tests). nForce4 works best in dual-channel modes, 1xDDR2-667 is a tad slower, while single-channel DDR2-533 always lags noticeably behind.

Here is the situation in games: Doom 3 and FarCry in low modes prefer nForce4 SLI IE (a couple percents), three best nForce4 modes are on a par again. nForce4+1xDDR2-533 is on a par with i925XE. But at higher resolutions (and consequently complex graphics quality) both Intel chipsets suddenly shoot forward, retaining their advantage over NVIDIA even under the heaviest conditions. As memory speed and other factors have no effect on the performance, we can only assume that manufacturers of these chipsets (or these motherboard models) optimize PCIEx16 operations on the driver level (or on the level of hidden BIOS parameters) in a different way.

However, Unreal Tournament 2004 does not show this effect, all contenders are very close to each other here, but the general alignment of forces remains.

Conclusions

I repeat that today's conclusions about the advantages of DDR2-667 over DDR2-533 can be applied only to processors with 1066MHz FSB. This issue is not relevant for the majority of simple Pentium 4 models (and obviously the future Pentium D), DDR2-533 with good timings will be more than enough for them.

Going back to the results of today's tests, I want to note the following points:

1. On the whole, nForce4 SLI IE turns out faster than i925XE, though this difference is almost impossible to make out in practice. The i945P model that we tested has demonstrated relatively low performance, but we are not inclined to take it for the final result: let's wait for the official announcement of the chipset, probably new drivers, and certainly new BIOS versions.
2. The chipsets are no slower with DR2-667 memory than with DDR2-533 at the same timings, but no faster either (under dual channel conditions) — fractions of a single percent are insignificant. Considering that the overwhelming majority of DDR2-533 memory offers 4-4-4 timings, there is absolutely no sense in striving for DDR2-667 with extra low latencies. But we render homage to Corsair diligence and capacities — its CM2X512A-5400UL modules will long be famous as the fastest memory available.

Sergei Pikalov (peek@ixbt.com) May 17, 2005.