| Summary : |
DDR-II :
Ttiming Vs frequencies
Now we will study the influence of DDR2 timings. We know that in certain cases the timings are more important than the memory frequencies and vice versa. However, as we saw for the increase in frequency, it seems that DDR2 reacts very differently than DDR1. For this reason we wanted to measure the affect of the timings on DDR2. For that, we will carry out various tests:
- Pentium 4 'EE' 3.20 GHz - 16*200 - 3:4 - DDR-II 533 - CAS 5-5-5-15
- Pentium 4 'EE' 3.20 GHz - 16*200 - 3:4 - DDR-II 533 - CAS 4-4-4-12
- Pentium 4 'EE' 3.20 GHz - 16*200 - 3:4 - DDR-II 533 - CAS 3-3-3-8
- Pentium 4 'EE' 3.25 GHz - 13*250 - 3:4 - DDR-II 667 - CAS 5-5-5-15
- Pentium 4 'EE' 3.25 GHz - 13*250 - 3:4 - DDR-II 667 - CAS 4-4-4-12
- Pentium 4 'EE' 3.20 GHz - 12*266 - 1:1 - DDR-II 533 - CAS 3-3-3-8
- Pentium 4 'EE' 3.20 GHz - 12*266 - 1:1 - DDR-II 533 - CAS 5-5-5-15
- Pentium 4 'EE' 3.20 GHz - 12*266 - 1:1 - DDR-II 533 - CAS 4-4-4-12
As you can see, aside from the 50 MHz of variance on 3.2 GHz in 667 MHz mode, all the modes are clearly comparable. Now then, we will compare the 800FSB/DDR2-533 mode with 250FSB/DDR2-667. In the interest of being comprehensive (which is desirable), we also included the mode 1:1 at 266 MHz. We start with a fast outline under Sandra:
This graph will enable us to note the baseline figures of the timing/frequency evolution, which we’ll verify next. First, it is clear that in asynchronous mode, the influence of the timings is much smaller than in synchronous mode. Indeed, while going from CAS 5-5-5-15 to 4-4-4-12 (or 4-4-4-12 to 3-3-3-8) brings a gain of only 2% in asynchronous mode, it delivers a boost of more than 5% in synchronous mode. However, at identical timings, DDR2-667 remains faster than the DDR1-533 synchro mode, except at CAS 3 timings, which currently are unattainable using DDR2-667. In any case, the advantage over the traditional 800FSB/DDR2-533 mode is obvious. Now let us see how is the latency, measured with mBench:
If the ranking seems identical to the throughputs obtained, it is interesting to note the influence of timings at low frequencies. Indeed, the DDR2-533 CAS 3-3-3-8 mode appears only 4% slower than DDR2-667 at CAS 5-5-5-15. In any case, DDR2-667 at CAS 3-3-3-8 would only make it possible to go down under 70 ns, which remains high compared to DDR1. And it is clearly the weakness of the DDR2. However, the latter has a strong point: its maximum bandwidth. So let’s look at the scores obtained under wstream:
The DDR2-667 mode again takes the lead in the race here and we again see the importance of timings in synchronous mode. Indeed, the jumps in performance when the timings decrease remain above those of the asynchronous modes. For now, let’s forget synthetic benchmarks and move to the practical benchmarks. We start with a game, good old UT2003:
In UT2003, it is clearly DDR2-667 that finishes in the lead since even at 5-5-5-12 latencies it outpaces DDR2-533 at 3-3-3-8, and this, even at 1:1 with a 266 MHz FSB! Proof that synthetic benchmarks don’t always give a realistic picture of the gain seen in real applications. Now we’ll try to confirm the results witnessed here with some 3D rendering tests under Kribi:
Here still, we see a ranking strictly identical to the test under UT2003 where DDR2-667 takes first. Furthermore, it’s obvious that the low timings have a large influence on overall performance. To finish, an encoding test with Windows Media Encoder 9:
Here also, the ranking is identical and you can see that to obtain the best performance, it is necessary to either remain synchronous with low timings, or to go up higher, in DDR2-667 mode, with timings that are still reasonable. However, it is noted that even at 5-5-5-15 settings the asynchronous mode does better than 1066FSB at 1:1.