| Summary : |
DDR-II :
Synchro (1:1) Vs Desynchro (3:4)
FSB800 Vs FSB1066
Before moving on to DDR2 vs DDR1 tests and the influence of the timings, it seemed interesting to us to give a progress report on the performance of the asynchronous mode of the i915/925. Indeed, we know that 800FSB Pentium 4 requires a 6.4 GB/s bandwidth and that this bandwidth is reached by two DDR1-400 channels. So what is the interest in offering 8.5 GB/s? In theory, none. In practice, the effectiveness of the memory controller report not being 100%, the gain can be worthwhile. It is what we will check today by measuring the gains from the asynchronous mode. So, we will test the following configurations:
- FSB 800 Mode
- Pentium 4 3.2 GHz -> FSB : 800 - RAM : DDR-II 400 - Ratio 1:1 - Coef : 16*800 - 3.2 GHz
- Pentium 4 3.2 GHz -> FSB : 800 - RAM : DDR-II 533 - Ratio
3:4 - Coef : 16*800 - 3.2 GHz
- FSB 1066 Mode
- Pentium 4 3.2 GHz -> FSB : 1066 - RAM : DDR-II 533 - Ratio 1:1 - Coef : 12*1066 - 3.2 GHz
- Pentium 4 3.2 GHz -> FSB : 1066 - RAM : DDR-II 354 - Ratio 3:4 - Coef : 12*1066 - 3.2 GHz
As you can see, the final frequency of 3.2 GHz is preserved and everything is comparable, with the exception of the DDR2-333 mode, which becomes DDR2-354 in 1066FSB mode. It is besides for this reason that Corsair offers not PC2-5300, but PC2-5400 which is specified for 1066FSB mode 3:4. A good idea. The timings are also fixed at 4-4-4-12. The motherboard used is a DFI 925x-t2 and the graphics board an ATI Radeon X600 clocked to 350/350. Just for kicks, let’s see the results under Sandra 2004:
Before commenting on this benchmark, we should explain why we obtain 4.65 GB/s here in 800FSB mode - DDR2-533 whereas we would have expected 4.9 GB/s. The reason is simple, to be able to go below a 14x multiplying coefficient, we had to choose a Pentium 4 EE based on the Northwood/2M core. This one produces worse results on the Sandra memory test. Why? It’s curious, but we noted that by adding a little more than 5% to all the scores, we arrived at the results of the P4 Prescott. On this test, we saw that moving from DDR2-400 to DDR2-533 offers a performance gain that is not negligible. Of course, it doesn’t equal the move to a 1066FSB, but even in this mode, the gain brought by DDR2-667 is considerable. Let us continue with the mBench "Latency" test:
As you can see, latency decreases sharply with the FSB, which is logical. However, we noted that, it is for 800FSB mode or 1066FSB, the 3:4 asynchronous mode has some considerable advantages over the 1:1 synchronous mode. Now let’s look at the pure throughput that we measured with wstream:
The results are here clearly in favor of the asynchronous mode, which offers a performance gain of approximately 15% compared to the 1:1 synchronous mode. So you advance from 1.95 GB/s to 2.4 GB/s in 800FSB mode, and from 2.7 GB/s to 3.1 GB/s in 1066FSB mode. The low efficiency of the memory controller of the i925X (approximately 75%) probably explains this performance gain. Let’s continue with a practical application: Kribibench:
In the same way, one sees here a linear gain between the modes. Let’s finish with UT2003, in BotMatch mode:
Same thing in UT2003 where the performance gains between the various memory modes appears good. Moreover, the X600 that we use didn’t make the performance increases more visible, but a more powerful graphics board makes it possible to post a gain of approximately 5% between the 1:1 mode and the 3:4 mode.