Memory ratios and timings analized

I spent about an hour collecting data for this. None of the tests were repeated multiple times because the ranges were large enough that they didnt look like they were in the magin of error. All of the tests were done in 3dmark01 with the details set to the lowest settings and the card at artifact free speeds and with all of my normal background programs running.

Heres the score for my settings that I (used to) use down in my sig: 21441

(In case I change my sig before someone reads this here's a copy/paste of it: Main rig:3.0C(ES) @ 3.7GHz | 246FSB, PAT enabled | 1GB of BH5 @ 246MHz/2-2-2-6 | R9800 256M @ 435/378 | Audigy 2 ZS | WD 80GB 7200RPM 8MB cache )

I then set the ratio to 5:4 and tested that: 20956

Theres not nearly as large of a performance drop as people claim, and as I will demonstrate that loss is easily recovered by the additional clock speeds you can acheive, but first the 2-2-2-5 VS the 2-2-2-6 timings issue.

2-2-2-5: 21441
2-2-2-6: 21443

I would reccomend you put it at 2-2-2-6 because of the potential overclocking gains and no loss in performance, I would imagine that 2 points is within the margin of error.

Now after I discovered that 5:4 wasnt such a huge loss I took the CPU and FSB up much higher.

I found my max stable speed at 275FSB and 220MHz memory, and now my cpu can run at 3855MHz. The new score for this setup was: 21723

I tried 295FSB, but that crashed in the middle of 3dmark even though my cpu speed was only 3840MHz, I'm mobo limited right now.

At 290FSB and 3770MHz I scored: 21577

So in conclusion, I say to drop your ratios and crank the cpu as much as it can go with a high FSB. Only set your ram to 1:1 if you can do it without loseing more than 100MHz on the cpu speed or 10MHz on the FSB.

Nice analization.

Sophisticated said:

Nice analization.

Click to expand...

ROFL

Anyway, very informative

That's pretty much what we've been saying in another thread.
21577 / 21433 = 1.0067 or about 6 tenths of one percent faster. Certainly within the margin of error for 3DMark. Regardless, if you get 100 fps one way, you'll get 100.6 the other. Not really a deal breaker

At some point we went past the threshold where memory bandwidth limits current games. I don't know where this is. Maybe around 180mHz? Now it seems that video power and raw cpu speeds are the kings, as long as you stay past that (as yet undetermined) threshold. I've run a couple games and 3DMark with different memory timings and 1:1, 5:4 and 3:2 ratios. Big difference in SiSoft Sandra memory bandwidth tests, very small difference in 3DMark, and virtually nothing in games. I was, in fact, slightly faster at 3:2 with 2-2-2-6 timings than at 1:1 at 3-4-4-8.

If you're doing a lot of work with large picture files or if you OC as a hobby, then the 1:1 ratio is important. If you OC for gaming and you've got a P4 2.4 or higher, get some nice PC3200 or 2700 with tight timings and run at 3:2. If you've got a dual memory channel motherboard, get 2 sticks of memory and make sure it's double sided. It's too bad, I learned my lesson a little too late this time.

Gary

well, acctually it went from 21443 to 21723, but it doesnt change the meaning of your post.

I was doing this with my ES chip so I could help people figure out what cpu to get based on their ram. A 2.4C with a 5:4 ratio will do just as good as a 3.0C with a 1:1 ratio. There was acctually alot of information I was attempting to find out for my friend who was trying to deceide what new CPU/ram/mobo combo to get, but it seemed relevent for alot of people so I posted it here.

Nice info.

Can you try some 2-3-3-6 and 2.5-3-3-6 tests as well?

those arnt exactly controversial settings. If its important to you test it yourself, theres no reason why you cant.

Well, its basically the difference between Winbond CH5 and BH5. I've never had sticks that could do tRCD 2, and probably never will. Eh, no worries.

you can drop the speeds down to like 190MHz and test it. I have CH5 sticks that do 2-2-2-6 all the way up to 198MHz.