RAM Dividers - A Comparison for Core2 / DDR2 / 975X

[Angry_Games]

Ok, once again I've been spending time trying to get a definitive answer, so that we can just link everyone here like we do to the Tight Timings Vs Loose Timings vs 1T vs 2T thread and the 939 vs AM2 thread.

 

We've all been wondering whether or not memory dividers are still the performance killer on these new Core2Duo processor platforms. Remembering back to the old days of Pentium4 and AthlonXP, we knew that dropping to dividers would either corrupt your bios (heh, NForce2!) or give you a bit of performance drop (Pentium4), and dropping to 2T Command Rate would do the same thing (again, worse performance hits on AthlonXP than on P4, but the P4 on a divider @ 2T was extremely poor compared to 1T @ 1:1 on the same Mhz overclock).

 

I've heard various claims that DDR800 is much better than DDR533 divider, memory bandwidth is as or more important than cpu mhz. I've heard others (and spouted it myself) that again, like the Athlon64, none of that is important compared to raw cpu mhz.

 

Now is the time to figure that out, and on a valid platform, though a bit...unorthodox as you'll see when you see the specs:

 

DFI Infinity 975X/G

2x1GB Corsair Dominator (XMS) DDR2-800 @ 4-4-4-12

2x Leadtek 7950GT 256MB in SLI (yes, SLI with the 84.56mod2 driver...it's fantastic)

Intel Core2Duo E6400 2MB L2

Seagate 7200.10 250GB 16mb cache SATA II hdd

NEC 3550A DVD-RW

OCZ GamerXstream 700w psu

Vigor Gaming ATX case (it's the CM Stacker clone!)

 

 

I'm in the beginning stages, but I was just fiddling with the RAM divider while I am at 3.0Ghz (375x8) and remembered that I wanted to test all of this.

 

 

 

(Any of you that have read my rants before already know all of this story and where it will lead to, so I'll just cut to the point heh...if you want to know more of the story of Core2 and why we hate Intel blah blah, search this forum or google and find out).

 

 

 

So fast forward to now using the rig in my sig, a 3Ghz Core2 with X1900XT Crossfires. I was the first to admit once I'd gotten my hands on one that they blew away my fastest Athlon64. Later I would find out that even when I finally reached 3120Mhz on an AM2 X2 4800+, it just couldn't compete with 3Ghz Core2.

 

But the question always persisted...does it make a difference if I use memory dividers considering the Core2 still uses what we would almost consider an antiquated Memory Controller Hub...or Northbridge sitting between cpu and RAM...a well known bottleneck (for both data and overclocking lol)?

 

Especially since I'll be using 2 7950GT's in SLI mode (yes, it works on non-Nforce SLI chipsets if you use the search button), we should determine how important that memory bus really is in most of the things we will use our machines (and I'll get into some more in-depth testing as time goes on, this is just a preliminary bit so I can get my ideas down on the forum, then shut my yap-hole and test test test). At 1600x1200 resolution with all the eye candy maxed out, we'll be able to hopefully see what dividers are going to do to performance (as those high resolutions with everything maxed really moves a lot of data...makes the machine stress extremely hard)

 

 

 

 

So, here's my testing methodology, both of what I am doing right now, and will be doing.

 

It's a bit out of order, but I will arrange it better as more comes in lol.

 

375x8 @ DDR400 divider, 4-4-4-12, 2T

375x8 @ DDR533 divider, 4-4-4-12, 2T

(since I'm already there doing it lol)

 

266x8 @ DDR800 divider, 4-4-4-12, 2T

266x8 @ DDR667 divider, 4-4-4-12, 2T

266x8 @ DDR533 divider, 4-4-4-12, 2T

266x8 @ DDR400 divider, 4-4-4-12, 2T

(note this is stock speed for E6400, and 2T is the default for this memory/chipset, but I will get into 1T stuff soon enough. Right now, this is about dividers)

 

3d Mark 2001 SE @ default settings

3d Mark 2001 SE @ 1600x1200, 4x Anti-Aliasing, Triple-Buffering (max settings basically)

 

3d Mark 2003 @ default settings

3d Mark 2003 @ 1600x1200, 4x AA, 16x Anisotropic Filtering (AF) (max settings)

 

3d Mark 2005 @ default settings

3d Mark 2005 @ 1600x1200, 4x AA, 16x AF (max)

 

3d Mark 2006 @ default settings (remember, it is 1280x1024 in 2006, not 1024x768!)

3d Mark 2006 @ 1600x1200, 16x AF Note here: Nvidia 79xx series cards are either incapable of running HDR + AA at the same time, or the driver hasn't been written yet...but you cannot use HDR + AA with these drivers and I couldn't get it to work on the NF590 SLI board with 93.71 drivers either so I'm doubtful that it is the driver)

 

 

Like the other threads, I will provide the chart (though I am tempted to learn how to make those pretty bar-graph charts like all the review sites do!) with various scores and you can read over 'em and if you want to argue 'em, you can test yourself and show everyone what you come up with.

 

Like those other threads also, we'll keep the flaming and the I-Know-It-All ish stuff to a minimum.

 

Core2 is new to me as it is to all of you, and I'm not experienced enough yet to declare I'm right and you stink like I am about Athlon64 and AthlonXP/Pentium4 stuff lol.

 

But this is what this thread will be about.

 

enough yapping dude! give us numbers!

 

ok heh

 

 

 

more as i get to it

[Angry_Games]

placeholder thread

 

 

oh, and because I know this will come up, I'll answer it before anyone asks:

 

So Angry, after all this railing about 3d synthetic benchmarks, why are you subjecting us to them?

 

 

short answer: because I haven't installed any of the games just yet and already started banging out these benches which are the first things to get installed when I am putting machines into the Stock Speed and Overclocking Database sections

 

 

long answer: all that above, but at the same time, these are familiar benchmarks that everyone recognizes, everyone knows their own scores, everyone knows probably some outrageous scores...we mostly know where these benchmarks will score on a certain combination of hardware. That's how many all of us have seen in the last few years.

 

Though they are synthetic, they are comparing directly with the same test multiple hardware combinations, settings, etc. They are valid for that reason alone, yet at the same time, each newer benchmark incorporates the latest technology into it's testing, so now that every card has SM3.0, it can be tested as it is an important visual aspect of games, and games are what we be playing the most on these shiny new computers.

 

I never look at say 3d Mark 2006 and try to correlate how Call of Duty 2 will run based on 3d Mark's score. That's just not realistic. It sorta is...but not really. Though you can at least correlate that if you can pull 30fps in say the arctic snow test steadily, you can probably play Call of Duty 2 without much problem...but narrowing down performance more than that is not my bag (which is why I test the real game haha...but just haven't yet!)

 

But again, it's a familiar set of numbers that we all recognize, and it does a fairly good job of comparing machine to machine. So I continue to use them (I also use them as stability test for the Database entries, but when I run them then, they are boring lol...much more exciting to compare everything for a reason....a reason better than saying "well it works"!).

 

ok, that was just in case anyone wanted to take me to task for railing against 3d mark and Aquamark etc then they are the first benches I post up

 

CoD2, FEAR, all that stuff that I have I'll get to, don't worry...it's the holiday man, my fat arse can barely get around the lab!

 

 

 

 

 

 

EDIT: oh and ps, I will be doing some tight timings vs loose timings and 1T vs 2T all that jazz no sweat!

[farksy]

Just a small suggestion from me.

 

I've just moved to Core 2 Duo, and i got to say, most will be very curious on the performance difference.

 

Instead of having all at Cl4 ... try

 

266x8 @ DDR800 divider, 5-5-5-15, 2T

266x8 @ DDR667 divider, 4-4-4-12, 2T

266x8 @ DDR533 divider, 4-3-3-8, 2T

266x8 @ DDR400 divider, 3-3-3-8, 2T

 

Or something like that. 1T vs 2T wise, 2T allows you to clock much better, but since the memory controller is not on-die, the performance gain is still negligible.

 

Nevertheless, still one of those comparisons everyone would like to see. Thumbs up, and i'll be waiting for more results

[steelballzz]

once again nice testing

 

with my E6600 i noticed that for pi 1 m it works very good with 1:1 with tight timings....like 3-3-3 .....but for pi 32 the fsb is critical...even with 5-5-5 i get better results at fsb 533

 

waiting for more results

 

thnxs

[WintersFury]

Interesting to see tha tthe numbers change less and less as we get to the newer 3dMarks. It'll be interesting to see what results using real games has.

[Chaos_2k]

@ Happy_Games:

 

375x8 @ DDR400 divider, 4-4-4-12, 2T

375x8 @ DDR533 divider, 4-4-4-12, 2T

(since I'm already there doing it lol)

 

266x8 @ DDR800 divider, 4-4-4-12, 2T

266x8 @ DDR667 divider, 4-4-4-12, 2T

266x8 @ DDR533 divider, 4-4-4-12, 2T

266x8 @ DDR400 divider, 4-4-4-12, 2T

 

Im a total noob when it comes to Intel but, am i right if i think that 266*8 @ DDR800 Div. means 1064mhz on the mem ? then the other settings would have less mhz on the ram without changing the timings? Is this true?

 

In order to dont confuse anyone here i think it would be a good idea to add the "real ram-mhz" too.

 

Sorry for bad english

[Angry_Games]

Just a small suggestion from me.

 

I've just moved to Core 2 Duo, and i got to say, most will be very curious on the performance difference.

 

Instead of having all at Cl4 ... try

 

 

 

Or something like that. 1T vs 2T wise, 2T allows you to clock much better, but since the memory controller is not on-die, the performance gain is still negligible.

 

Nevertheless, still one of those comparisons everyone would like to see. Thumbs up, and i'll be waiting for more results

|

|

|

V

EDIT: oh and ps, I will be doing some tight timings vs loose timings and 1T vs 2T all that jazz no sweat!

[Angry_Games]

once again nice testing

 

with my E6600 i noticed that for pi 1 m it works very good with 1:1 with tight timings....like 3-3-3 .....but for pi 32 the fsb is critical...even with 5-5-5 i get better results at fsb 533

 

waiting for more results

 

thnxs

 

super pi is not a test I will be using because it is a completely pointless test other than saying you have the largest e-penis on the internet

 

this is absolutely not about that

 

this is about real programs in real conditions...I'll be testing games, 3d benchmarks, video editing/rendering (Sony Vegas 6.0d), etc. As much "real" stuff as I can

[Angry_Games]

@ Happy_Games:

 

 

 

Im a total noob when it comes to Intel but, am i right if i think that 266*8 @ DDR800 Div. means 1064mhz on the mem ? then the other settings would have less mhz on the ram without changing the timings? Is this true?

 

In order to dont confuse anyone here i think it would be a good idea to add the "real ram-mhz" too.

 

Sorry for bad english

 

at DDR800, the defautl divider, the memory is running at 800Mhz, but the cpu's FSB is running at 1066Mhz

 

it's normally shown as 4:3 (4 being the 1066FSB, 3 being the 800mhz RAM)

 

once you start moving down the divider list, yes the memory frequency does get lower, and when overclocking, the FSB does go higher.

 

for instance, right now I'm @ 3.0Ghz overclock, which is 375x8.

 

I'm on the 533 divider, which is actually means I am running 1:1

 

ie: the Core2 is a 1066Mhz FSB cpu right out of the box. Since we know everything is DDR, that means it is a true 533Mhz bus (double-data-rated...it would be doubled, or 1066Mhz).

 

so since I am up to 375x8, I am running 1:1, which means my RAM speed is doubled (again, using DDR), to 750Mhz.

 

Keep in mind however that my FSB numbers change in accordance to the cpu's stock FSB.

 

so at stock speed, I'm at 1066FSB....but at the overclock of 375x8, my FSB is now 1503Mhz.

 

here's where the math gets tricky for those of you not familiar...

 

 

Intel cpu's, since Pentium4 days, have had what is called a "Quad-Pumped Bus".

 

This means that the core2 cpu, is really only a 266Mhz FSB. But since we use DDR RAM, it's doubled to 533.

 

Normally you can imagine a clock cycle in a processor to be like a real clock's pendulum....it swings all the way left, then all the way.

 

This is one complete clock cycle (left to right back to original left)

 

Processors from the old days were able to execute instructions at the end of one complete clock cycle. So a 200Mhz cpu using 200Mhz RAM would only be an effective 200Mhz total in communication.

 

AthlonXP's and such moved on to DDR RAM.

 

DDR RAM was really only SDRAM. The difference is that processors were now able to execute instructions on each "edge" of the clock cycle.

 

For the 'edge' again imagine it like a real clock's pendulum, swinging back and forth.

 

 

Now Intel, in order to compete with the superior AthlonXP, decided to keep that same 200Mhz bus, but they designed the cpu to execute TWICE per clock edge.

 

so, on SDRAM, if you got 200Mhz FSB and 200Mhz RAM, you were at 200Mhz

 

on DDR SDRAM, if you had 200FSB, and 200Mhz RAM, you were at 400Mhz, because the cpu could execute instructions at each edge of the clock cycle, thereby "doubling" the data being manipulated (but we chose to explain it as a speed increase to make it easier to understand).

 

 

Intel's Pentium4, PentiumD, and Core2, they can execute TWICE at each edge of the clock cycle.

 

so if you have a 200Mhz FSB, 200Mhz RAM, you are at 800Mhz. Again, if you execute two instructions on each edge of the clock cycle, you have effectively 'quadrupled' the amount of data being manipulated.

 

The only real change now is the Intel cpu's are 266Mhz FSB

 

still quad-pumped bus

 

so

 

266x4 = 1066Mhz.

 

 

on the Core2 right now, the FSB is still 1066Mhz, but it is only accessing the memory controller @ 800Mhz (because honestly, it really only needs 533Mhz RAM)

 

so the RAM is running at 800Mhz, while the cpu is running at 1066 (reality only 533 but we'll let Intel have it their way for sake of explanation)

 

this is why you get the odd divider ratio of 4:3, 2:3, 5:4, etc. It means that the cpu and the memory are not running in synch (1:1, or 533/533).

 

But the point of all this is to test just like I did for Athlon64...how important are memory timings? How important is 1T vs 2T? How important is memory bandwidth vs cpu mhz?

 

So far, these processors are identical to the Athlon64 in the sense that CPU Mhz seems to be the king....but we'll see as I get more testing

[green-man]

happy as ive said before, nice 1 ure a legend. hope my feeble attempt of a ram testing thread pushed u to doing this.lol.

as u might of noticed ive been scratching my head for weeks, trying to work it out.

the best speed ive managed to get out my ram, thats sis sandra ram speed bench. is at 412fsb/ 533div 4-4-4-12 which is just above 800MHz ram, at that setting im almost touching 8GHz in sis ram bench. but obvouslly thats when im pushing my setup bigtime.

i can on a slacker divider push my ram well above its rated 1000MHz but the speed is nothing compaired to what i get above fsb400 no matter what divider/ram timings??

is it just sis is'nt mesuring what i think its mesuring sort of like u said about superpi being a . contest, coz their taking speed from the cpu not just the ram thats what the funny results are about?

[Angry_Games]

happy as ive said before, nice 1 ure a legend. hope my feeble attempt of a ram testing thread pushed u to doing this.lol.

as u might of noticed ive been scratching my head for weeks, trying to work it out.

the best speed ive managed to get out my ram, thats sis sandra ram speed bench. is at 412fsb/ 533div 4-4-4-12 which is just above 800MHz ram, at that setting im almost touching 8GHz in sis ram bench. but obvouslly thats when im pushing my setup bigtime.

i can on a slacker divider push my ram well above its rated 1000MHz but the speed is nothing compaired to what i get above fsb400 no matter what divider/ram timings??

is it just sis is'nt mesuring what i think its mesuring sort of like u said about superpi being a . contest, coz their taking speed from the cpu not just the ram thats what the funny results are about?

 

the point of it is though is to decide if, like the Athlon64, the cpu Mhz is the key to performance over everything, and what, if any, do dividers, Command Rates, and timings play in cpu'z overall mhz performance gains.

 

the Athlon64 we basically see that timings, command rates, and dividers are moot points, it's all about raw cpu Mhz, so the higher the cpu mhz, the more performance (ie: breaking the stigma of dividers/cpc/timings being so totally important as to be avoided)

[Chaos_2k]

Hey Angry keep us posted.

 

I've found something simmilar in the meanwhile (sorry it's on a german forum but words like FSB and MHZ you all do know:angel: ).

The author has do some some serious testing and whats the result:

CPU mhz is King, followed by ram mhz followed by timings.

 

 

http://www.forumdeluxx.de/forum/showthread.php?t=293050