Performance

Synthetic Memory Subsystem Benchmarks

Since the launching of the i925/i915 platform brought us a completely new memory subsystem based on DDR2 SDRAM, we should first of all study its performance in synthetic benchmarks. At first we will use ScienceMark 2.0 utility with pretty rich opportunities for extensive memory testing. First of all we measured the memory bandwidth and memory latencies obtained on platforms with Pentium 4 processors. We will compare the results for the new DDr2 SDRAM with those for the regular DDR400 SDRAM. The table below contains our measurements taken from Socket 478 and LGA775 platforms running with Pentium 4 CPUs based on different processor cores but working at the same clock frequencies of 3.4GHz. Moreover, we also added here the results for Socket 939 Athlon 64, Socket 940 Athlon 64 FX and Socket 754 Athlon 64. In order to make it a fair comparison we tested AMD64 processor with 2.2GHz core clock.

The obtained results show that as we have expected, DDR2 memory features higher practical latency than DDR400 SDRAM. However, despite its higher theoretical peak bandwidth, DDR2-533 cannot boast higher practical bandwidth than the regular DDR400 SDRAM. The thing is that the 8.5GB/sec bandwidth provided by the dual-channel DDR2-533 cannot be used to the full extent by the today’s Pentium 4 processors with 800MHz bus, because this processor bus features lower bandwidth of only 6.4GB/sec. This way Pentium 4 processor will be able to enjoy all the advantages of the DDR2 memory only when they acquire 1066MHz Quad Pumped Bus support. There is not so much waiting left, actually: Intel is planning to announce these CPUs some time in Q3 this year already.

Now let’s take a look at the results our testing participants showed in the memory subsystem tests of the SiSoftware Sandra 2004 package using Stream algorithm to measure the practical bandwidth of the memory subsystem:

Note that the systems working with DDR2-533 SDRAM are again showing lower results than systems working with DDR400 SDRAM. As a result we should only state that Intel’s desire to equip its platforms with the new memory type is more like a good move for the future, which advantages we will feel some time later. So far, the memory subsystem built with DDR2 SDRAM cannot please us with high performance in synthetic benchmarks. At the same time, I should say that it is still too early to conclude whether the use of DDR2 SDRAM slows the contemporary platforms down compared with the regular DDR SDRAM. All in all, the performance depends on the algorithms involved into work with the memory subsystem that is why systems with DDR2 SDRAM can still work faster than those with DDR400 SDRAM in certain specific applications.

I would also like to draw your attention to the fact that DDR2 SDRAM reveals its potential much better when it works with the Prescott based CPUs. It can be probably explained by the hardware and software data prefetch algorithms, which underwent serious modifications in these CPUs. Anyway, you’d better keep this peculiarity in mind. Later on we will see if this tendency is also true in real applications, or if it reveals itself only in synthetic tests.