Choggs396

EDIT (7-24): I've included some 3rd party tests on page 2. Thanks.

Well there has been a lot of discussion as of late regarding Multipliers, FSB:RAM Dividers, and performance thereof. Some have argued that running a divider will increase performance by allowing higher RAM frequencies. No one will argue that higher Memory frequencies mean higher bandwidth (and thus, higher performance) as a general rule. But I have argued that if you are capable of obtaining a Front Side Bus at the same speed as your RAM (well, half, speaking in terms of DDR of course), that you should keep a 1:1 ratio. In my opinion, running at a ratio that is "out of sync" will deter performance.

Now I'd like to make a few things clear here. I will state again that higher bandwidth is almost always better. But I am strictly arguing that in the case of memory speeds being exactly the same, a ratio of 1:1 will trump the performance of a ratio of anything else.

Of course, the only way to test this would be to change the FSB. I still recieve arguments that lowering the FSB, increasing the Multiplier (to effectively have the same CPU speed), and running with a memory divider to achieve the same RAM speed will provide us with better overall performance.

For example:

I say that my system, being set at this:
e6600 - 400FSB X 8 Multi = 3200MHz, memory running at 1:1 FSB:RAM (800MHz)

is faster than:
e6600 - 356FSB X 9 Multi = 3204MHz, memory running at 8:9 FSB:RAM (800MHz)

Blacklight115 and I had a very interesting discussion about all this at this thread that he started.

Well I decided to test this theory out using 3DMark06.

I set up my system using all the same settings as follows, except for the ones you see above:

RAM timings 4-4-4-12 1T

VCore @ 1.35V BIOS. Stable overclock using OCCT 2 Hours.

8800GTX @ stock speeds for all tests (575MHz Core / 1800MHz Mem)

Ambient Temps of about 18-19C

CPU Idle Temps ~29C

------------

I decided to run 3DMark06 for 4 passes using the 2 setups.


***** TEST RESULTS *****


CASE #1:

E6600 @ 3200MHz, 400FSB x 8, 1:1 FSB:RAM ratio (sync mode), memory @ 800MHz

Photo of the BIOS:


Screenshot of CPU-Z Confirming Specs:


And a screenshot of Speedfan for Temps:


3DMark06 Results:

10902 3DMarks
10922 3DMarks
10919 3DMarks
10861 3DMarks

Totaling 43,604 3Dmarks .... or 10901 Average 3DMark06 Score.


************

CASE #2:

E6600 @ 3204MHz, 356FSB x 9, 8:9 FSB:RAM ratio, memory @ 800MHz

Photo of the BIOS:


Screenshot of CPU-Z Confirming Specs:


And a screenshot of Speedfan for Temps:


3DMark06 Results:

10865 3DMarks
10888 3DMarks
10839 3DMarks
10852 3DMarks

Totaling 43,444 3Dmarks ... or 10861 Average 3DMark06 Score.

*************

Testing Conclusions:

So, CASE #1, using the 1:1 FSB:RAM Ratio, averaged:
10901 3DMarks

Opposed to CASE #2, using an 8:9 FSB:RAM Ratio, averaged:
10861 3DMarks

The FSB and Multiplier were changed, while both cases were tested @ 3.2GHz and @ 800MHz RAM

*************

While these numbers aren't greatly different by any means, they were done with 4 tests each to find a good average. And the FSB was changed only slightly for a reason - I wanted to see a noticeable change over the two even at small frequency changes while running a divider.

While I admit this doesn't prove beyond a reasonable doubt that my original speculations are correct, it does hold a little weight to my argument that the divider does reduce performance, if even a little. This is arguable but I've provided some evidence to back it up.

I understand that the exact reason why the divider may be causing a performance decrease is purely speculation at this point. The performance decrease could possibly be attributed somehow to there being asynchronous transfer of data occuring between the RAM and the FSB which causes some kind of slight "bottleneck", so to speak, or slowdown of some kind. This is speculation, however.

Let me just say again that there is no exact proof here and by no means am I implying such. I just believe there is some credibility to these tests.

Thanks for taking the time to read this. By all means, comment away... please no flames though, it took me hours to put this together.

And special thanks to blacklight115 for encouraging me to create this test situation.

And also: If you see any errors or miscalculations in my numbers, please inform me and I'll fix them.

PS: Whew! I need a break. Be back later.

__________________
Je pense, donc je suis
- René Descartes

Cooler Master ATCS 840 Owner's Club

Overclock.net Professionalism Initiative

Witchfire

Impressive testing, Choggs. Rep+

__________________
Some people live on this earth to do great things... Unfortunately, most are here simply because it's illegal to kill them.

12,277 in 3DMark06 on a mostly secondhand budget rig

Ellipsis

Nicely done Choggs. Did you run SuperPi while you were at it?

Do you plan on running more tests in the near future? Maybe drop down to a multi of 7, max out the FSB and try a 1:1 divider on the RAM with looser timings, say 5-5-5-15?

I'm craving input. I've got my sig rig tore down for a cable management job or I'd be joining you in all the fun.

This is what I come here for. Not the Romper Room antics of late... trust me.

+Rep

coldchillen

good job.. read this earlier, but didnt have time to reply.. yeah, id like to see more tests run in the future.. itd be really helpfull for fellow newbies like myself, trust me

__________________
im sorry, brenda. i dont know any other way..
Intel Multiplier Lower Than It Should Be? READ THIS!!
Multipliers Changing? READ THIS!!

Choggs396

You're seriously going to kill me, but I'm actually flying out to Florida tomorrow morning for 5 days. That's one reason why I didn't wait to do the initial testing, because it was either tonight or wait 5 days.

And another big reason why I wanted to finish it beforehand, was because I could see comments like this if I didn't: "Oohh luk he prolly ran teh test an faild so hes gunna just say hez leeving. lol noob FTL, u sukzorz".

This is really bad timing I know, but I'm sure I'll have lots of fun. I'd ask my girlfriend to run the tests, but hmmm... no. I told her she could only touch my Athlon XP rig (even though I love my Barton, too )

I'll have a laptop available to me, so I'll prob make some posts if I'm not too drunk.

But rest assured, when I do make it back I plan on this:

SuperPi tests. (really sorry I didn't do them already),

More 3DMark06 tests:
E6600 @ 2.4GHz (400X6), 1:1 FSB:RAM ( Memory @ 800MHz )
-vs-
E6600 @ 2.4GHz (266x9), 2:3 FSB:RAM ( Memory @ 800MHz )

^ That should be a really good one. I actually think I should have done it initially.

Well I have to get up in 3 hours. lol. Talk to ya soon.

Thanks for the enthusiasm!

EDIT: And I have definitely contributed to some "Romper Room antics". Sorry. I've been trying to improve.

__________________
Je pense, donc je suis
- René Descartes

Cooler Master ATCS 840 Owner's Club

Overclock.net Professionalism Initiative

Ellipsis

I think you took that the wrong way. The "Romper Room Antics" crack wasn't directed at you, personally. I was talking about the general direction the forum, as a whole, has taken lately. But I digress. Enjoy your trip. Looking forward to your future benchmarking results.

MrBungle

your theory simply cannot be proved. changing the multiplier can put you in a higher strap, which will raise latency. lowering your FSB will slow your system down. SuperPI is not a test of overall system performance. you are trying to compare apples to oranges.

the closest my simple mind can come to a semi accurate experiment would be to run your FSB at 266 with memory at DDR 533, then running your FSB at 266 (same multi) and memory at DDR 1066. of course, that is a useless experiment because we all know what the outcome will be.

alexisd

Choggs396 what part of Florida you goin?Im in the southeast if you need a quick stop.
And what he is trying is change, to lower multy to see if is any gain in performance.

__________________
6700 ES @ 3.50http://valid.x86-secret.com/show_oc?id=182370
6850@4.04=http://valid.x86-secret.com/show_oc.php?id=222356

MrBungle

i believe he's trying to demonstrate that running your system at a 1:1 ratio will yield better performance than a 1:2 ratio, which someone else claimed was ideal (in a different thread. i forget who). while i admire his tenacity, i think he's wrong on this one.

luckii

of course higher fsb would yield better performance but only slightly. @ 3.2, i could use either 400x8 w/ 1.4325 or 356x9 w/ 1.4 vcore. so i opted for the less vcore =)