Picking new memory

[BradyT88]

What do you think the fastest DDR2 memory is? I'm preferably looking for a 2x2GB kit that can do about 1080 @ 4-4-4-12-12-2T or 1440 @ 5-5-5-15-20-2T or @ 6-6-6-18-24-2T. I've been told on another forum that my best bet would be DDR2 1200 5-5-5-15 G.Skill Trident or Pi.

I have my PhII 925 OCed to 3.78GHz (270x14) so with a ref clock of 270 setting my memory divider to DDR2 800 gives me 1080 and setting it to DDR2 1066 gives me 1440.

Another point assuming that I can pull of the 1080 @ 4-4-4-12 and the 1440 @ 6-6-6-18, which would give me the best performance as far as gaming and why?

I've been doing a lot of reading up on timings and I read that when accessing a row of memory, the CPU sends the message to the memory and memory will delay "tCL" clocks until returning the info and any other info needed on that same row is sent every clock without delay. I don't know exactly how data is organized on memory, but I would assume it is fairly sequential. So is it possible that having the faster clock will more than make up for the longer delay in the long run.

For example, DDR2 800 cas 4 has a clock of 2.5 ns (nano seconds), and a cas latency of 10 ns, while DDR2 1066 cas 6 has a clock of 1.875 ns and a cas latency of 11.25 ns.

The 800 is going to beat the 1066 at sending the first bit of data, but by the fourth bit of data (assuming all 4 are on the same row) the 1066 wins, 17.5 ns vs 16.875 ns. Well say you are opening a program or loading a level on a video game and it was reading 4 bits of data on each row from 100 rows. You would save 62.5 ns (not factoring tRCD and tRP).

Looking at this I would say a faster clock is always better (with the exception of really loose timings compared to tight timings) and if you are getting more than 4 bits of data per row this would be even more of an advantage.

Any thoughts on this?

 

[BradyT88]

No thoughts huh? I went ahead and bought the GSkill Trident DDR2 1200.

 

[BradyT88]

I got my GSkill in the mail yesterday and learned that I made a small under sight. my tRCD and tRP cannot be set lower than 5 on my board. My tRAS cannot be set lower than 15 and when I tried 1066 4-5-5-15-20-2T, my computer wouldn't even post. The JEDEC on this memory is 400 5-5-5-15-24 @ 1.8v and there is an EPP that is 533 5-5-5-5-15-48-2T @ 1.8v. 1066 5-5-5-15-20-2T passed a memory test and has been running fine. I tried 1200 5-5-5-15-20-2T and there was no post. 1093 5-5-5-15-20-2T passed a memtest, but I got a random BSOD about 3 minutes into a movie.

 

[ekoostik]

Did you look at articles comparing performance differences. You mentioned differences of "17.5 ns vs 16.875 ns" - but when you get to RAM that is that close in SPEC the actual performance difference - outisde of synthetic benchmarks - are barely noticeable if at all.

 

[BradyT88]

Well after lots of playing around, I got these sticks to 1080 ( could possibly go higher, my cpu was only stable up 270x14 so I didn't try higher) @ 5-5-5-5-16-2T @ 2.1v. The Jedec on them is 800 5-5-5-15-20-2T @ 1.8v and there is an epp, 1066 5-5-5-15-43-2T @ 1.8v.

One question though, how many passes of memtest is considered stable? These sticks on currently on pass 4 (about 2.5 hours) and no errors so far.

A problem though, these have successfully made it through 3 passes of memtest, but they fail a Prime95 blendtest, (cpu underclocked to 270x8, HT and NB both at 270x4 and the mem is all the way up)...is memtest not foolproof or is my issue somewhere in my chipset or IMC or what?

 

[ekoostik]

Let 'em run through memtest overnight. That'll give you a good degree of confidence.

Memtest is a very thorough test that the memory chips do not have deffects. It doesn't test your system under load. So if you pass memtest, you can be confident that the RAM does not have underlying defects, but you still have to use the other tests to check for overall system stability with your OCed settings.

 

[BradyT88]

I see. I ran it for 7 hours and it went through 9 full passes.

 

[BradyT88]

So is there anything about DDR2 memory preferring an HT and NB at 1000 MHz rather than 2000. I have my ref clock set to 250, my cpu multi at x14 (3.5GHz) @ 1.5v (so I know it's stable), my HT and NB both set at x8 (2000MHz) and my memory set on the 1:2 divider, which puts it at 1000MHz @ 6-6-6-18-26-2T @ 2.1v. I had the memory isolated and it ran stable for 6 hours. With everything else OCed, it failed prime blend about an hour in. I decided to drop the HT and NB multipliers to x4 (1000MHz) and it has been running stable for 10 hours now.

EDIT: Well I had some success. It was 24 hours prime blend stable. I turned the ref clock up to 260 and am trying again. Once I find a peak I will try to raise the HT/NB multipliers and tighten the timings.