In spring we reviewed functionality of the new Intel chipset series — 945/955. Even though we had already published the first reviews of motherboards on these chipsets by that time, we took care not to speak of the performance of the new products on the whole. According to the first results, i945/955 chipsets not only failed to outperform their predecessors (i915/925), but were actually slower. Having assumed that the problem was in early BIOS versions of the motherboards on the new chipsets, we decided to wait for some time and carry out performance tests only when we were sure of stability. It's nearly six months already — quite a sufficient term in all respects for the manufacturers to fine-tune their products. So today we are going to compare i955X with i925XE and in the nearest future we shall publish a large roundup of motherboards on the new series of chipsets.
A few more words on functionality
We decided to omit the i945 results, as their difference from the i955X is not too large. And considering the spread of results depending on a model, its BIOS version, and test error, this difference disappears completely. You can consult our i915/925 roundup to look up the ball park, but we can say that an average motherboard on i955X is approximately on a par with the fastest i945 representatives. Reviewed together, i945/955 chipsets differ from their predecessors by 1066 MHz bus support. But as only two uber-expensive Pentium 4 Extreme Edition processors feature this FSB, we can dismiss this advantage over i915/925. Potential support for SLI mode is not an advantage to common users either. And of course, SATA-II support in ICH7/R does not provide performance advantage over SATA/+AHCI in ICH6/R, as hard drives still haven't learnt to read data faster than 150 MHz/s even from a buffer.
The main reason for launching i945/955 was their support for dual core processors — it's still their main trump card. This feature makes the chipsets unique. The chipset choice (among Intel products) for Pentium D or Pentium Extreme Edition processors is not in question. Among other characteristics, there seems to be only one factor speaking for i945/955 — they support DDR2-667 memory. It's clear that 667 is greater than 533 (and 400), so this product is obviously attractive to an inexperienced user. But experienced users are aware that the 800 MHz FSB throughput of the Pentium 4 processor (maximum frequency for all modern processors from Intel, except for the two above mentioned processors) is just 6400 MB/s, which corresponds to the throughput of dual-channel DDR-400 (in case of 1066 MHz FSB, it will be dual-channel DDR2-533). Theoretically, excessive memory bandwidth comes in handy during intensive data exchange with other devices (besides a processor), e.g. a 3D accelerator, mandatorily accompanied by intensive data exchange with a processor. In practice, such a situation looks exotic, but we shall still hold back our comments till the real tests.
However, there is one factor, which may make higher-performance memory modules more attractive. It has to do with the second vital characteristic of a memory system — access latency. Indeed, a manufacturer specifies (in SPD) and a user can change (in BIOS Setup) memory timings, which determine latencies during typical low-level memory operations: CAS# Latency, RAS# to CAS# Delay, and Row Precharge (remember that most modern modules ignore TRAS). The typical DDR2 latency is 15 ns, which gives us 3-3-3 for DDR2-400 mode, 4-4-4 for DDR2-533 mode, and 5-5-5 for DDR2-667 mode. Memory modules with such timings are quite typical. At the same time, a number of manufacturers offer (for a higher price, of course) modules with reduced timings — for example, popular DDR2-533 with 3-3-3 and DDR2-667 with 4-4-4 (as DDR2 cannot have the key timing CAS# Latency below 3, "advanced" DDR2-400 modules are not available). You can actually find higher-performance modules, if you pay through your nose (an example is not far to seek). But we shall not seriously analyze the results of still exotic products, if you please. Thus, one of the objectives of this article is to find out when modern chipsets for Intel processors really need supporting memory faster than DDR2-400, and whether they need this support at all.
We have chosen a motherboard from Intel as the i955X representative
in these tests. With the latest BIOS versions, it caught up with the
other models on this chipset and its lag from the leader (ASUS
P5WD2 Premium; it does not take part in these tests for technical
reasons) became purely nominal. The i925 choice was easier: Gigabyte
8AENXP-D proved itself the best model in our i925XE
roundup. And the only difference between the two i925 modifications
is the support for 1066 MHz FSB. We have taken "improved" memory modules
(according to our classification), which allow to startup at 533 MHz
and 667 MHz with their timings lower than typical values by one.
Of course, they were also used to emulate "normal" modules. Well,
let's proceed to the results.
The chipsets and all memory modes under review are equal in terms of their memory read rates (measured in RightMark Memory Analyzer), the maximum spread of results does not exceed 3%. It's not surprising, as we have already mentioned above that the main bottleneck is FSB. A little subsidence of the maximum read rate (with prefetch) is typical of 2×DDR2-400. In fact, the maximum rate is a tad higher than the theoretical one (6400 MB/s) due to the enlarged cache (L2) of the processor.
The situation with the memory write rate is much more interesting, though the non-temporal store mode always yields the same results (Gigabyte motherboard's results are higher by one per cent precisely - that's how much its FSB frequency is raised). The real write rate grows noticeably (by 12—17%) when we upgrade from 2×DDR2-400 to 2×DDR2-533. It reacts tangibly (+5—10%) to reduced DDR2-533 timings. i955X and i925X[E] chipsets demonstrate similar results here. Further frequency growth and reduction of timings in DDR2-667 memory yield practically no effect, though we can still squeeze a couple of per cents.
Access latency is directly determined by memory timings, so we naturally see the same results in all operating modes with the same typical latency. DDR2-533@3-3-3 is a tad faster than DDR2-667@4-4-4, as it should be. i925XE slightly outscores i955X in pseudo-random access latency (which shows memory access latency proper), but the random access (characterized by often D-TLB misses of the processor) is noticeably better in i955X, so the memory controller of the new chipset has been probably overhauled. You may download (RAR, 37 KB) RMMA test results for several other memory modes (with various timings) and evaluate their performance effect on your own.
We'll start our tests in real applications with archivers, as the most sensitive to memory performance. 7-Zip and WinRAR demonstrate an absolutely identical picture, but the results of the latter archiver are more illustrious, as 7-Zip operates with much larger memory volumes, transferring which smoothes the difference over. One cannot but notice that archiving performance is determined not only by the memory latency, but also by write rates, so the 2×DDR2-400 mode is a tad slower than modes with higher memory frequencies and the same (in absolute values) latencies. i925XE is outperformed by its successor, as its latency of random access to memory pages is worse. On the whole, the difference between the configurations is not large, about 8% within a chipset.
Video encoding performance, measured by our open method, nearly doesn't depend on memory performance, we have seen it many times. There are some small differences (not exceeding a couple per cents), they comply with the tendencies, we already noted. But even the FSB frequency, raised by one per cent (and thus the CPU clock), of the Gigabyte motherboard yields much gain.
Though SPECviewperf results are of no interest for the comparison of desktop platforms, we constantly use this test, as it allows to detect interesting facts about the graphics bus and provides the marginal memory mode, which loads a processor and a video card, heavily exchanging data between each other. We published the results of three out of eight sub-tests, which demonstrate the most interesting dependences. 3dsmax-03 is nearly irresponsive to changing a chipset and a memory mode. But it demonstrates a huge performance drop (~40%) in 2×DDR2-400 mode on i925XE. proe-03 is more sensitive to a memory write rate, slow DDR2-400 and DDR2-533 modes in this test lag behind by up to 6%. But the performance drop (15—20%) demonstrated by i925XE/2×DDR2-400 eclipses everything. ugs-04 is obviously responsive to memory latency, demonstrating a significant 9% spread between different modes on i955X. But it's covered by the 17% lag of the i925XE/2×DDR2-400 (i925XE fairs quite well in other modes).
The conclusion is obvious: overhauled memory controller and/or graphics bus controller (PCI Express x16) in the i955X fixes the problem, which existed in the predecessors of this chipset in this marginal mode: memory bandwidth and FSB bandwidth match, there is additional need in data transfer for peripheral devices. But let's see how this effect shows itself in real 3D applications — games.
We can see nothing unexpected in Doom 3. Lower-performance modes are indeed slower. The maximal difference is 4%, i925XE is absolutely no worse than i955X, all testbeds demonstrate identical results at 1600×1200 with the highest image quality.
The situation in Unreal Tournament 2004 is similar (the difference is below 3%), but the unfortunate i925XE/2×DDR2-400 mode actually slumps in 1600×1200 — by 6%.
The i925XE@2×DDR2-400 lag effect is the most noticeable in FarCry, already in 1280×1024 mode with high image quality. The slump may reach 30%, while the other modes differ from each other by no more than 5%.
If you need a dual-core processor from Intel and a motherboard on a chipset from the same manufacturer, your choice is limited to the i955X and the i945 series. But considering the similar other functionality of i915/925 and i945/955 chipsets, DDR2-667 support in the latter looks their only advantage for single-core processors. However, in actual fact this support provides no performance advantages in standard modes. A little performance gain can be achieved by installing DDR2-533 and DDR2-667 memory modules with reduced timings. Laying aside the question of expedience of buying these memory modules, we should note that DDR2-533@3-3-3 is faster than [equally priced and available] DDR2-667@4-4-4 in all real applications. Thus, buying a motherboard on i945/955 makes sense only if you are going to install really elite DDR2-667 modules with their key timings not exceeding 3.
In the process of testing we suddenly found out another argument for the new Intel chipsets: as we found out, when the graphics and the processor buses are simultaneously loaded to maximum, the i925XE (we can expect the same from the i925X and the i915 series) starts operating very slowly in 2×DDR2-400 memory mode. This effect is always present in synthetic SPECviewperf as well as in the heaviest modes in some games. However, DDR2-533 even with nominal timings solves this problem just as well. Besides, it looks more reasonable. Thus, dual-channel DDR2-400 memory is quite sufficient in the majority of typical situations. Enthusiasts may pay more money for DDR2-533 with reduced timings, while rich fans may pay through their nose for DDR2-667 with super low timings (a motherboard on i945/955 is justified only in the last case). Overclocking FSB is another matter, but the general information provided in this article is evidently enough to predict that such overclocked system will need more-or-less fast memory.
Write a comment below. No registration needed!