Jammer

Hey guys,

Decided to do a bit of testing to see how memory speed and latencies affect performance. In particular I wanted to see how games are affected by RAM speed and latencies. I chose 3DMark06 as I see no point in running older 3DMark benchies except for e-penis value.

Setup:
Pentium D 930 @ 4.0GHz (266 * 15) Stock Volts
DFI Infinity 975X
2GB G.Skill PC6400 HZ (Micron D9 IC's) Stock Volts
XFX 7900GT VIVO (550/1630 Factory OC)

Results:
DDR2 533 4-4-4-5 1:1 = 5415 3DMarks
SM2: 2193
SM3: 2209
CPU: 1989

DDR2 800 4-4-4-5 2:3 = 5412 3DMarks
SM2: 2193
SM3: 2199
CPU: 2001

DDR2 533 3-3-3-5 1:1 = 5417 3DMarks
SM2: 2191
SM3: 2203
CPU: 2010

DDR2 800 4-4-3-5 2:3 = 5412 3DMarks
SM2: 2192
SM3: 2201
CPU: 2001

DDR2 667 3-3-3-5 4:5 = 5403 3DMarks
SM2: 2192
SM3: 2204
CPU: 1975

I don't have screenies so you guys will just have to take my word for it. But interestingly enuff, in games at least, there is no advantage in running the RAM faster than ur FSB, in fact it even decreases performance slightly. So it looks like the old 1:1 rule is still correct (for games ayway). This is good news as you can run tighter latencies on ur RAM.

I might even test RAM at 667 to see if this provides any performance benefit... Stay tuned.

EDIT: Added DDR2 667 results. Very Interesting!

Jammer

Running PCMark05 benchies now... different story all together guys.

Jammer

PCMark05 results here:

DDR2 533 3-3-3-5 1:1
PCMark05 Score: 7191

DDR2 667 3-3-3-5 4:5
PCMark05 Score: 7328

DDR2 800 4-4-3-5 2:3
PCMark05 Score: 7306

As you can see, in PCMark05 the results are slightly different. Unfortunately I don't have the registered version of PCMark05 so I cannot pin point where the differences in performance are but it still tells us that:

- For gaming 1:1 is fastest.
- For general PC use, 1:1 with a bit of an "overspill" (to quote the all wise and mighty RodneyJM) buffer seems to yield the best results.

Now I know that this has probably been discussed many times before but after hearing soo many conflicting optinions, I had to test it for myself.

Comments welcome all...

eva2000

3dmark06 isn't mem/bandwidth dependent much so timings won't help much

RodneyJM

Jammer, you don't need the registered version to see if there is any difference with ram timings and bandwidth. But it does help, however as you can see from your results 4:5 is the best. That does require more Vdimm though than 1:1 with tight timings regardless of DDR2 sticks on market at moment.

There is no "overspill" with 1:1, 1mhz goes in to the MCH and another goes out...etc...synchronously. As little time lag as possible on Intel platform.

Jammer

Sorry when I say 1:1 with a bit of overspill i mean 4:5. 5:5 = 1:1 so 4:5 = 1:1 with a bit of overspill. If that makes sense?

eva: Yes I see this... NOW!

Unless of course you're running G.Skill HZ's. All my tests were run with 1.9V VDIMM regardless of timings. 1.9 in BIOS is about 1.94 actual.

RodneyJM

ahh, no.

3-3-3-5 is not tight as in the true meaning of the word "tight" eg.
Go as far as the bios can take you with this DDR2 option for timings.
Can your board to 3-2-2-4-2 ? how do you know its 1.94v ? use a DMM ?

Jammer

It needs 2.2V but yes it can! Not using a DMM just going by what the BIOS and SmartGuardian tell me.