3-3-3-2-1T DDR2 is worth it!

[Simps]

Hello!

I am very pleased with the 3-3-3-2-1T DDR2 times...
This is from an old 680i chipset.
Only 420MHz RAM, on a 3.6GHz C2Q / 450 FSB

Everest MEM


I bet, on the X38 (that can handle +600FSB), a DDR2 running @ only ~480MHz 3-3-3-2-1T will perform just like (or better) then the 700MHz CL5 that I have seen here...

Can anyone, with that kind of setup, try it?

[]'s
Simps

[binormalkilla]

I'm pretty sure that 1T isn't possible on any Intel chipset with DDR2.....however I could be wrong.
Intel chipsets benefit from higher frequencies much more than tighter timings (when it comes to bandwidth)
I'd try it.....but I know that I won't POST and I'll have to swap out my value RAM so I can set the voltage, then reboot and install my Ballistix

[MaSell]

680i loves tight timings
965P, P35, X38 prefer MHz for sure

btw some p35 asuses can run with 1T IIRC

[KTE]

I tried 3-2-2-2-1 2T at 450MHz on P35 but I've only seen ASUS offer the option of CR which I didn't have and it wasn't worth it to me more than running 450FSB 1:1 3-3-3-3 2T was. My latency was lower than what you're getting there though IIRC and on this 790FX setup, I've tred 4-4-4-4 1T at 480MHz @ 1.9V and it wasn't worth it for any useful application. I won't try less because it'll need more volts which I don't want to give it.

When I tried 450FSB 562MHz 4-4-4-4 CR2 it was actually higher than what you're getting there: http://img85.imageshack.us/img85/408...5624444jn5.png

And 675MHz 5-5-5-25 CR2 was even higher: http://img85.imageshack.us/img85/8849/hshshspc3.png

Never bothered to run any more tests on bandwidth or drop timings and tweak TBH.

[binormalkilla]

My latency is much lower than that....IIRC it was 56.4ns or so. I know I've gottena s low as 55ns on this X38 board. I run 1066 5-5-5-15 35 with the rest set to auto. I don't tweak my memory much, as I really don't run SuperPi.

[saaya]

kte, i think thats cause you used higher fsbs with higher mem clocks for cas4 and cas5 right?

on 680 i think 333 beats 444 and 555 if you use the same fsb and the max speed each timing is stable at.

the problem on intel chipsets is:
1. no command rate 1t, its called cmd 1t but its not really command rate 1t
2. very very few dividers and no dividers below 1:1, so you can run high fsb low mem clock with tight timings...

[Simps]

This is interesting...
680i / C2Q@3.6 / 450FSB


DDR2@420Mhz 3 3 3 2 1T



DDR2@573Mhz 4 4 4 12 2T



DDR2@619Mhz 5 5 5 15 2T



Same results... but the latency read from Everest, on the 3-3-3-2-1T is better...


[]'s
Simps

[RPGWiZaRD]

That's a bit high latency for the CL5 config.

Seems really like 680i likes tighter timings. At least these comparisions speaks so. I haven't tried CL3 timings on my P35 board but at CL4-4-4-8 at or slightly above 500MHz I get worse results than 563MHz 5-5-4-10.

Check my numbers at my 24/7 config:

[Simps]

@ KTE

Nice bandwidth man!
P35 delivers more memory bandwidth then 680i...

573Mhz 4 4 4 12 2T on my 680i = 9488 read
562MHz 4-4-4-4 2T on your P35 = 10025 read

I still believe, that something arround 480Mhz and 3-3-3-2-1T would perform similar to +700MHz CL5...

[]'s
Simps

[Simps]

@ RPGWiZaRD

You are running on a higher CPU clock and FSB...
The numbers can't be compared... Try doing 8x450 (3.6GHz), and then post the Everest readings...

And also, P35 has some better memory bandwidth / latency over the old 680i...

[]'s
Simps

[RPGWiZaRD]

@ RPGWiZaRD

You are running on a higher CPU clock and FSB...
The numbers can't be compared... Try doing 8x450 (3.6GHz), and then post the Everest readings...

And also, P35 has some better memory bandwidth / latency over the old 680i...

[]'s
Simps

I know all that thanks. Was just pointing out the latency difference, a 150MHz higher clock won't show that much difference.

But sure I could do a comparision at 450 too... I'd wanna know what kinda subtimings you use in that case if you want to make accurate comparisions. I've tweaked every subtimings in MemSet so it shows the best results according to Everest & SuperPi when focusing on stability too (lowering some might show next to no difference so I keep it at the higher number to compensate for stability).

[saaya]

interesting results guys!
thx for sharing

simps, did you tweak the subtimings as well?
the 333 was with 1t while the 444 and 555 was with 2t right?

[Bl4ck^!]

yep 1T only on nForce chipsets , my ABIT AB9PRO had 1T and 2T selectable but choosing 1T made whole setup unstable no matter what memory , also if there was some stability there where no usable benefits of it.

[Simps]

@ Saaya

Yes.
333 with 1T
444 and 555 with 2T

The subtimes are not tweaked...

[]'s
Simps

[KTE]

I still believe, that something arround 480Mhz and 3-3-3-2-1T would perform similar to +700MHz CL5...

On P35 it won't if you have low timings/subtimings on the +600 run, but it will if you have high timings/subtimings. CAS is only 1 of many parts to your bandwidth/latency. Mine is affected a lot by CR, tRC, tWTR and PL, but the last three more than with CAS or CR.

That 675 5-5-5-25 T2 run had PL6 and low subtimings.

This run of 676 5-5-5-15 T2 has PL7 and higher subtimings: http://img220.imageshack.us/img220/7782/675bj4.png

= ~900MB lower bandwidth

[Pyr0]

these are the settings i've been trying (untweaked runs):
450FSB, DDR2 900, 4-4-4-10, 1T


450FSB, DDR2 1200, 5-5-5-15, 2T


450FSB, DDR2 1200, 5-4-4-12, 2T


450FSB, DDR2 1300, 5-5-5-15, 2T


not a huge difference
i believe the differences we are seeing in bandwidth scores on similar chipsets may be due to the difference in quad and dual core

[Simps]

@ Pyr0

OMG! Very good man!
Can you try 3-3-3 with 1T?

I liked that 450MHz 4-4-4-10 1T
I will try that on my 680i...

[]'s
Simps

[RPGWiZaRD]

It sure is a great difference between 1T and 2T on 680i, wow that's some awesome numbers for such low clocks. On Intel board you'd have to have to run really nice ram speeds in order to achieve 10k+ read at 3.6GHz.

[Pyr0]

just done a couple of small 3-3-3-4 runs at 1T for you:

400FSB, DDR2 800, 3-3-3-4, 1T


450FSB, DDR2 800, 3-3-3-4, 1T


450FSB, DDR2 840, 3-3-3-4, 1T


450FSB, DDR2 900, 3-3-3-4, 1T (definitely not for 24/7 use)


unfortunately my sticks aren't so great at Cas3 clocking and need more voltage for similar bandwidth and latency as a higher cas.
because of this i tend to use Cas 4 and 5 with lower volts, hopefully for a longer lifetime

[KTE]

The 680i boards are better with bandwidth/latency. That was one of their specialities: memory. Intel with 1:1 is always very good at any FSB.

When I ran 450FSB 1:1 450 4-4-4-4 T2 I had the same SuperPi 1M time as 540 4-4-4-4 T2 as with 562.5 4-4-4-4 T2 and 630 4-4-4-4 T2. 562.5 was actually fastest by 0.1s. I've never really bothered with bandwidth though so not tested many settings I've ran.

[Simps]

@ Pyr0

Man, that is crazy good results!
Congrats... I can't do that on my 680i...

[]'s
Simps

[jcool]

Yeah 1T performance really is nice on the 680i.
However, you can run DDR1000 @ CAS 4 with same or fewer voltage... and atleast on my old P965 it was quite fast



Couldn't run 1T higher then ~700 or so but that was just using 2,2V which is perfectly ok for 24/7 use
I will be having my X38 here soon so I'll check back.

About Intel and 1T, the Abit IP35 Pro can run 1T up to ~1000 (so I'm hoping the X38 QuadGT can do it as well). However, the performance gain isn't as high as with nvidia.

[Simps]

Man, this result from pyro @ 450MHz DDR2 is insane.



Can't believe this

[]'s
Simps

[Tony]

There are also nice latency sweet spots on the MCH of 680i,if you know where to look you can gain some quite stunning performance linked sync 1T.

I remember seeing 6400MB/s+ in memtest at ddr900ish 1T while testing the DFI 680i...could get no where near that 2t with high clocks but I had to be at a certain FSB where the MCH was probably at a huge clock also

[The0men]

The reason (my guess) why pyro has such good latency is due to tighter timings across the board. I can get about 44.6ns with DDR1000 CAS4 and 1T.

680i certainly are easy to make boot with 1t thats for sure, but you can seriously reduce timings by drastically pulling back sub-timins, especially where you can more than halve them like tRC and tRFC.

Sometime you have to take from peter to give to paul, and occasionally even reducing these subtimings you can lose bandwidth.