Performance
Before we go to the systems performance in real applications, let’s check how fast the new chipsets work with the memory subsystem. So, let’s resort to the synthetic Cachemem test, which we have been using for a long time now to test the new sets of core logic.
| Memory read speed, MB/s | Memory write speed, MB/s | Memory copy speed, MB/s | Latency |
Pentium 4 3.0 | i875 | 3422.1 | 1114.1 | 2268.8 | 248 |
Pentium 4 3.0 | i875 | 3255.2 | 1060.6 | 1910.5 | 278 |
Pentium 4 3.0 | i875 | 3038.9 | 888 | 1607 | 288 |
Pentium 4 3.06 | i875 | 2794.1 | 955.6 | 1794.2 | 302 |
Pentium 4 3.06 | i875 | 2852.9 | 903.9 | 1647.1 | 334 |
Pentium 4 3.0 | i865 | 2639.1 | 1113.8 | 2223.4 | 282 |
Pentium 4 3.0 | i865 | 2894.3 | 1021.3 | 1842.1 | 305 |
Pentium 4 3.0 | i865 | 2664.1 | 876.2 | 1586.3 | 316 |
Pentium 4 3.06 | i865 | 2796.5 | 949.2 | 1781.3 | 290 |
Pentium 4 3.06 | i865 | 2851.4 | 898.9 | 1628.6 | 316 |
Pentium 4 3.0 | i845 | 2740.1 | 1023.5 | 1507 | 261 |
Pentium 4 3.06 | i845 | 2281.4 | 865.1 | 1302.5 | 328 |
Actually, the results speak for themselves. Nevertheless, I would like to comment on a few things. Firstly, let’s check out how efficient PAT technology implemented in i875 proved. According to the obtained results, the use of this technology allows not only to reduce the latency when the memory is addressed, but also to speed up the reads from the memory. These two factors explain why i875 outperforms i865.
Also we would like to point out that the use of 800MHz bus really does tell on the work with the memory. In particular, even if you use 800MHz FSB and dual-channel DDR266 SDRAM, this combination will prove faster in most tests than a combination of 533MHz FSB and DDR333 despite the equal bandwidth of the bus between the processor and the memory.
The single-channel i845PE chipset demonstrated pretty nice results with 800MHz bus. Its trump is low latency during the work with system memory. This fact may affect its performance in real tasks a lot, so that it could appear not much slower than the dual-channel chipsets.

All in all, the results of the synthetic SiSoft Sandra 2003 benchmark measuring the practical bandwidth of the memory subsystem appear proportional to the theoretical bandwidth of the bus between the CPU and the memory in each particular case.
Bus frequency, MHz | Memory | Bandwidth of the bus between the CPU and the memory, GB/sec |
800 | Dual-channel DDR400 | 6.4 |
Dual-channel DDR333 | 5.3 | |
Dual-channel DDR266 | 4.2 | |
Single-channel DDR400 | 3.2 | |
Single-channel DDR333 | 2.7 | |
Single-channel DDR266 | 2.1 | |
533 | Dual-channel DDR333 | 4.3 |
Dual-channel DDR266 | 4.3 | |
Single-channel DDR333 | 2.7 | |
Single-channel DDR266 | 2.1 |
However, there are a few curious things here. I would like to stress that the practical bandwidth of the i875 chipset is always higher than the results shown by i865, which demonstrates how efficient the PAT technology is. It is also important that the FSB frequency exerts more influence over the path between the CPU and the memory than the memory frequency increase.

Another synthetic benchmark measuring the memory subsystem performance, which we used in our test session, is PCMark2002. In this test i875 also managed to beat i865, with the performance difference in some cases reaching 5-7%. As for the CPU, the processors with 800MHz bus managed to outpace those with 533MHz bus by 13-14%. We would also like to point out that the performance of i865 chipset appeared higher when we took Pentium 4 3.06GHz CPU with 533MHz FSB than that of i875, which indicates pretty clearly that PAT technology is most efficient when the FSB frequency equals 800MHz.

The processor test from the PCMark2002 test set showed that Pentium 4 3.06GHz with 533MHz Quad Pumped Bus was faster than Pentium 4 3.0GHz processor with 800MHz bus. This happened because the core clock frequency of the first CPU was higher: this benchmark almost doesn’t use the memory bus. This way it is hypothetically possible that newer processors will not be faster than their predecessors with 533MHz bus.
Now let’s move from synthetic tests to complex ones.



