The big memory ratio debate on 875/865

[Tony]

As you know when overclocking on Springdale or canterwood over 250ish fsb your going to have to use a downclock ratio on your ram.After reading Ed's article on Overclockers.com stating he though 3:2 was the way to go, I decided to do a few of my own tests.
Constants are:
motherboard=P4P800 bios 1008beta
FSB=272
Ram =TwinMos BH5
CPU=2.6C

I tested using 2-5-2-2 Turbo all tweaks on at 3:2 and 2-5-3-2 at 5:4 Turbo all tweaks on.

5:4(320 ratio)
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/5to4cpuz.JPG

3:2(266 ratio)
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/3to2cpuz.JPG

First off im goung to show some sandra,both Buffered and UNbuffered at each settings.
5:4 buffered........................................................................................................UNbuffered
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/5to4buff.JPG http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/5to4-unbuf.JPG
And 3:2 buffered........................................................................................................UNbuffered
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/3to2buffered2-5-2-2-turbo.JPG http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/3to2-unbuf-2-5-2-2-turbo.JPG
Whether you believe sandra's results or not you can clearly see 5:4 GIVES MORE bandwidth than 3:2.Moving on to Mbench, we can see near identicle access latency but again 5:4 shows much better bandwidth.
Mbench 5:4....................................................................................................................3:2
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/5to4mbench.JPGhttp://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/mbench%203to2-2-5-2-2-turbo.JPG

[Tony]

With all these theoretical tests i though a bit of 3dmark, winrar and superpi was in order.
5:4.................................................................................................................3:2
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/5to4-3d.JPGhttp://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/3to2-2-5-2-2-turbo.JPG

I took 3 runs at each memory ratio and published the fastest.In each case 5:4 beats 3:2.
Moving on to super pi the 1M test was just not long enough to show any diference so I moved to the 4M test instead.At 8M or greater im sure the diference in speed would be greater than we see here.
5:4................................................................................................................3:2
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/4Msuperpi5to4.JPGhttp://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/4Msuperpi3to2.JPG

Lastly some winrar, i decided to archieve XP SP1a with best compression with both NTFS options ticked in advanced to try to make the cpu and memory work harder.

5:4.....................................................................3:2
http://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/winrar5to4.JPGhttp://homepage.ntlworld.com/ad.leach/P4P800/memory%20ratio%20test/winrar3to2.JPG

Again we see 5:4 beating the 3:2 ratio and again im sure with a larger file the diference in speed would be larger still.

So whats my conclusion to all of this...i think Ed's got it wrong about 3:2 being faster ;)

[ThugsRook]

excellent post! it sure does get your point accross.

could it have something todo with the ratios themselves?
(not performing at their best yet)

* you need an unlocked chip to really get to the bottom of all this ~ and youve got the ram to do it too ;)

[Tony]

Thanks Thugs ;)

I think its just faster ram speed=faster results.Obviously latency has a big part to play and that is why i tried to keep the memory timings constant...but we all know RAS to CAS at 2 is a toughy once you get above ddr400 on these boards.

Regards my OCZ it has an spd flaw that stops it running async, this is why i was forced to use my old BH5 TwinMOS..still the results are valid as the same modules are used for both sets of ratio's.

[ThugsRook]

makes ya wonder if 1:1 would be even better :?:

of course youd need a 15-17x multiplier for that :beat:

[ThugsRook]

im gonna run some sandra ~ buffered and unbuffered (1:1 179fsb) just to see how close i get to your scores.

BRBIAB :wave:

[ThugsRook]

Bigtoe's results: buffered --- unbuffered

5:4 435ddr c2325 = 5921/5897 --- 2812/2829
3:2 362ddr c2225 = 5391/5378 --- 2599/2551

My results:

1:1 358ddr c2226 = 4562/4259 --- 2992/2838


* we are both using a P4P800 board & the same bios ver

(edit your numbers if i got them wrong BT)

[oldfart]

As you guys know already. I have a bit of a different opinion of mem ratios. 1:1 is fastest, the 5:4, then 3:2. This is true when running the same cas settings. No big secret. If you run a higher ratio with slower cas, its not the case. Some people will fall into this category.

The mem ratios are there for when you hit a wall with your FSB/mem speed. If your CPU/FSB speed can scale past your memory, drop to a ratio to get the CPU speed up. You can typically run a more aggressive cas setting to make up for much or all of it.

I really don't care much about synthetic mem benches since they have little bearing on actual performance in actual apps. This has been proven many times over.

Lets look at BTs 3DMark bench. 18062 (3:2) and 18111 (5:4). How much of a difference is that? It is a .27% difference. Just over 1/4 of 1%. It is not even close to anything you could ever notice.

Super PI, 1%.

Winzip is better with a 2.5% increase.

Run your CPU as fast as you can, use whatever mem ratio you need to get it there. Set your cas timings as good as you can and be happy. The ratios are not that big a deal in real apps. Synthetic mem benches do not apply.

[Tony]

If i didn't have to rip this rig to bits tomorrow i would do and mpeg2 encode at both settings and im sure that would clearly show what is faster.I posted at 272fsb as i thought this was close to what many would achieve with a 2.4 or 2.6C..i know most peeps are not made of money;)

My feelings remain, you need to balance access latency and bandwidth over raw cpu MHZ to get the best from your rig.

[oldfart]

Here are some MP3 encode tests using LAME.

For review, my system can hit one of two walls
1) Mem speed wall. My ram maxes out ~ DDR 426 with slow 2.5-3-3-7 timings. This equals 213 FSB max in this config.

2) CPU speed wall. My 2.6C can go to 257 FSB @ 1.6 Vcore.

This gives me a choice of

1) 213 FSB | 2.77 GHz | 1:1 | DDR426 2.5-3-3-7
2) 257 FSB | 3.34 GHz | 5:4 | DDR 412 2-2-2-5

LAME MP3 Encode
(lower is better)
257 FSB | 5:4 | 2-2-2-5
Encoded 11 files in 0:02:17

213 FSB | 1:1 | 2.5-3-3-7
Encoded 11 files in 0:02:45

213 FSB | 5:4 | 2-2-2-5
Encoded 11 files in 0:02:46

213 FSB | 3:2 | 2-2-2-5
Encoded 11 files in 0:02:46

Obviously, MP3 encode is all about CPU speed. 1:1, 5:4, 3:2 made no difference at all.

MPEG2 Video encode benches
TMPGEnc AVI -> SVCD CBR (lower is better)

257 FSB | 5:4 | 2-2-2-5
2:10

213 FSB | 1:1 | 2.5-3-3-7
2:35

213 FSB | 5:4 | 2-2-2-5
2:36

213 FSB | 3:2 | 2-2-2-5
2:40

Again, CPU speed is king. Also 1:1 slow timings and 5:4 fast timings are about the same. 3:2 is a little farther behind.

Some Q3 benches from another thread:

FSB | Ratio | cas settings | performance mode | Q3 FPS
257 FSB | 5:4 | 2-2-2-5 | disabled | 390.6
257 FSB | 5:4 | 2-2-2-5 | fast | 391.3

Next, 213 FSB since this is my max 1:1 FSB setting due to the limit of my ram. To get there I have to use slow timings:

FSB | Ratio | cas settings | performance mode | Q3 FPS
213 FSB | 1:1 | 2.5-3-3-7 | disabled | 335.8
213 FSB | 5:4 | 2-2-2-5 | disabled | 336.2
213 FSB | 5:4 | 2-2-2-5 | fast | 338.3

Didn't test 3:2 on that test.

All this talk about "MUST RUN 1:1" I still say is bunk. Every bench I've seen and run myself certainly says otherwise.

[Mikki]

bigtoe, excellent post and info! Very very interesting...;):D

oldfart, I know you like to downplay the minute differences we all get with these kind of tests, I see your point but I don't really agree. First, I know from tweaking my P4G8X and trying to get 15k in 3DMark2001 that getting a 50 point gain isn't as easy as it sounds. It may be a slight gain, but it is a gain. And second, if it's not worth tweaking your rig like this, then why are we here? :look:


1:1 is fastest, the 5:4, then 3:2 Is that a typo? :)

[RotorHead]

Mik, you know I’m going to have to agree with oldfart. I would rather have a 44Mhz gain in FSB then a 50 point gain in 3Dmark2001. ;) :p

Whichever setting is going to give me the highest overclock on the CPU will probably produce the highest 3DMark scores.

Having said that, I am definitely learning some good info on memory timings. :cool: http://www.bleedinedge.com/forum/images/icons/icon14.gif

RH

[ThugsRook]

if someone can get me a unlock C1 i will gladly purchase it and resolve this issue once and for all.

a single persons "comprimise" does not resolve this issue.

im willing to lay $ that 17x200 3400mhz 1:1 c22261 will be king over any 12x or 13x "comprimised system".

i personally think OCers got the shaft on these 800fsb systems.

:wave:

[Mikki]

Maybe I'm just confused at what everyone is trying to say, and with all the trouble we've been having with benchmarks lately, I believe it's very hard to prove anything. Real-world tests are good, but then if someone doesn't do a lot of large file transfers, what's the point in using something that shows them?

The only point I want to make is this (and I have no idea what, in any, relavence this has on the topic of discussion): All these little tests and tweaks we do here is to explore whatever we're reviewing. We do it to find the faults or advantages, to show whether something is good or not, and to tweak the system to get the best whatever possible. If an adjustment we make gives us a goal, or an improvement (or shows a loss), or even gives us a 5 point gain in a benchmark, then we did good IMHO. :)

[Tony]

I believe 250X13 for 3.25gig 1:1 ddr500 would be faster than 271X12 5:4.Not by much though as the cache on the chip is working much faster with a 12 multi at 271fsb than on the 13 at 250fsb.Short duration benchmarks are IMHO not good enough to show this..maybe a seti unit or an 8hr encode will.

[RotorHead]

Ok, I guess I’m looking at it from an overall end result instead of a specific gain. When you break it down a gain is a gain. :cool:

RH

[ThugsRook]

250fsb? ddr500 with aggressive cas and low vdimm? i doubt it.

450ddr tops for "typical" ram @ 2.75v max.

hehe ~ we really need to settle this debate :wave:

[Tony]

I can do 2-3-3-6 at ddr500 2.85v stable...the ram does 2-3-3-5 at ddr400 so the tests would be valid..just run TRAS at 6 at both speeds.Only problem is it won't do async :mad:

[oldfart]

I hope you guys don't take me the wrong way here. I'm not trying to downplay or discredit any of the terrific work that guys such a BT, Thugs, etc have done. Those that know me from AT know that I always speak my mind even if it goes against what the popular thought is.

FWIW, I did quite a bit of work on this stuff myself. I would like to think it is a valuable contribution and doesn't deserve to be downplayed either.

I'm just trying to point out that running a 5:4 ratio is not the end of the world as many seem to think. I've presented several common benches that show that 5:4 with fast timings is just as fast as 1:1 with slow timings. This can save people the expense of buying high end ram when they can get the same performance with what they have.

I've also shown it is not worth giving up CPU speed to run a 1:1 ratio. BTW, others around the net have also come up with that same conclusion.

I guess I'm a little different than some here. I'm not a benchmark junkie. I dont use my PC to run benches all day long. Unless there is a measurable performance increase in real applications, I don't get too impressed. 1%, 0.25% just doesn't do much for me.

Thugs, FWIW, we agree on most of this. Sandra mem benches are a joke. You will note I haven't posted a single one. Your feeling that 800 MHz FSB CPUs are overrated is well justified. A P4B rig can perform right there with a P4C rig, no doubt. What we differ is on the need to run 1:1 IF you have a P4C rig. My testing (and others) does not show this to be the case.

If there weren't people here with differing ideas, it would be pretty boring, wouldn't it?

Peace :-r

[Tony]

I do agree running 5:4 can be as good as 1:1..thing is if your only talking say 5fsb then 1:1 will win...if the diference is 20fsb or so then 5:4 is a must.

Also I really do wish this OCZ would run async for me as i would be able to show you just how good it is.PiLsY has a theory on it and im sure i will be able to get an indepth post from him on the subject.All i can say is there is a reason for 16 x 16MB(double sided) chips on each module that makes it come alive in dual channel mode.

Im also going to rebuild the SINXP 655 board i have with the 2.4b and see what the ram performs like on there.

[oldfart]

It would be nice to see that test. OCZ has come a long way! I am pretty limited in what I can test since I only have the one set of ram. I did what I could with what I have. XMS 3200C2 is also a pretty widely used memory type ,so I think it is fairly valid.

Again for me, I run 5:4 and 2-2-2-5 now. Is it even possible to get to DDR 514 2-2-2-6?? There are people running 2.4Cs and 2.6Cs with FSBs WAY higher than mine. They would need even better memory.

[Mikki]

Originally posted by oldfart
I hope you guys don't take me the wrong way here. I'm not trying to downplay or discredit any of the terrific work that guys such a BT, Thugs, etc have done. Those that know me from AT know that I always speak my mind even if it goes against what the popular thought is.

FWIW, I did quite a bit of work on this stuff myself. I would like to think it is a valuable contribution and doesn't deserve to be downplayed either.

I'm just trying to point out that running a 5:4 ratio is not the end of the world as many seem to think. I've presented several common benches that show that 5:4 with fast timings is just as fast as 1:1 with slow timings. This can save people the expense of buying high end ram when they can get the same performance with what they have.

I've also shown it is not worth giving up CPU speed to run a 1:1 ratio. BTW, others around the net have also come up with that same conclusion.

I guess I'm a little different than some here. I'm not a benchmark junkie. I dont use my PC to run benches all day long. Unless there is a measurable performance increase in real applications, I don't get too impressed. 1%, 0.25% just doesn't do much for me.

Thugs, FWIW, we agree on most of this. Sandra mem benches are a joke. You will note I haven't posted a single one. Your feeling that 800 MHz FSB CPUs are overrated is well justified. A P4B rig can perform right there with a P4C rig, no doubt. What we differ is on the need to run 1:1 IF you have a P4C rig. My testing (and others) does not show this to be the case.

If there weren't people here with differing ideas, it would be pretty boring, wouldn't it?

Peace :-r

It would get very boring indeed!!!! ;)

I agree with the above post, oldfart...even about the benchmarks generally. The only thing I will say is how are you going to measure a change/tweak in the bios to see if it does anything unless you use a benchmark? I know we could say that most of the bios tweaks make no noticeable difference, but tweaking is part of what we do and we need a way to measure it.

We could even take it further...if you could increase your oc by 10mhz would you do it? Sure you would. Would there be a noticeable difference? Nope. So why do it? :)

Please don't get me wrong, I'm not sticking up for benchmarks here, I'm totally with you guys, I think they're all cr@p and I think it's pretty sad that we don't have a single benchmark that scales the way it should and is 100% consistant. All I'm saying is what else are we going to use? Frustrating, isn't it? :rolleyes:

[ThugsRook]

what this thread really comes down to is "exactly which combo would provide the best results". (reguardless of budget)

using 3400mhz as an anchor, (and concidering current cpu & mem quality)
i personally think it will be in the 200-216fsb range, 1:1, & with fastest mem settings possible. (in talking c22261 with a 16-17x mplier here)

i cant wait to try and prove it one day :-)