How to run memory on Core2 platform...

[Octopuss]

So I read some article(s) stating that you get much more performance running memory in synchronous mode. Hum. I remember something about this several years ago. No idea about today though, with the rather fast memory and all...

The trouble is I cannot seem to find a proof for this statement. I got P35 board with 2x1GB of 1066MHz Kingston modules. FSB at 413MHz. So I did a simple test. Went down to 826MHz and even lowered some timings... and of course all tests showed big drop in performance. I got no idea what such articles meant by much more performance, but the standard ways of testing, at least in my eyes, prove otherwise.

What are your thoughts? Or someone even has experience with this?

[jcarle]

You may be talking of one of two things. Running your memory in a 1:1 configuration with your processor, that is matching the frequency clocks, or you may be talking about Dual Channel. If you're refering to the first, it's a matter of adjusting your processor's FSB to the highest possible value that can match what your memory supports. It's simple math once you understand how it works. Although you don't mention the FSB speed of your processor (or what processor you have, it goes something like this.

For a 1066 FSB CPU, it's base clock is determined by dividing by four (1066 / 4 = 266), which makes a 1066FSB CPU's base clock to be 266MHz.

Memory FBS is divided by two for DDR to determine it's base clock. So 800MHz DDR2 would have a base clock (800 / 2 = 400) of 400MHz.

To achieve a 1:1 ratio in this scenario, either the CPU base clock must be moved up to 400MHz giving it an effective FSB of 1600MHz or the memory must be slowed down to 266MHz, effectively making it run as DDR2-533.

In the case of your 1066MHz DDR2, it's pretty impossible to get CPUs to run at the base clock that your memory is running at (1066 / 2 = 533) as it would give an effective FSB (533 * 4 = 2132) of 2132MHz. The most realistic combination that is commonly used is running the CPU at 1600FSB ( 1600 / 4 = 400 ) with a base clock of 400MHz and running the memory at ( 400 * 2 = 800 ) 800Mhz.

[Witt3439]

I used Vista and the benchmark results there to check out my computer because I had the same question.

On the system I tested, it was running an E6300 on a 680i and had 2GB (2 x 1GB) of DDR2-1066 RAM (dual-channel). My memory score was at 5.6, so I switched off to 4 x 512Mb sticks of the same exact RAM (quad-channel now), and my RAM score was 5.9 and Vista actually felt a lot "peppier".

Same exact RAM (Corsair XMS2), speed, no overclocking. The only difference was the configuration and size of the modules.

In my opinion, the C2Ds like to be set up in a quad-channel configuration.

[Octopuss]

You may be talking of one of two things. Running your memory in a 1:1 configuration with your processor, that is matching the frequency clocks, or you may be talking about Dual Channel. If you're refering to the first, it's a matter of adjusting your processor's FSB to the highest possible value that can match what your memory supports. It's simple math once you understand how it works. Although you don't mention the FSB speed of your processor (or what processor you have, it goes something like this.

For a 1066 FSB CPU, it's base clock is determined by dividing by four (1066 / 4 = 266), which makes a 1066FSB CPU's base clock to be 266MHz.

Memory FBS is divided by two for DDR to determine it's base clock. So 800MHz DDR2 would have a base clock (800 / 2 = 400) of 400MHz.

To achieve a 1:1 ratio in this scenario, either the CPU base clock must be moved up to 400MHz giving it an effective FSB of 1600MHz or the memory must be slowed down to 266MHz, effectively making it run as DDR2-533.

In the case of your 1066MHz DDR2, it's pretty impossible to get CPUs to run at the base clock that your memory is running at (1066 / 2 = 533) as it would give an effective FSB (533 * 4 = 2132) of 2132MHz. The most realistic combination that is commonly used is running the CPU at 1600FSB ( 1600 / 4 = 400 ) with a base clock of 400MHz and running the memory at ( 400 * 2 = 800 ) 800Mhz.

I absolutely meant the first. Synchro with FSB.

Btw. You are talking about some prehistorical C2Ds The 65nm E6750 I got has default FSB of 333MHz.

I understand how this thing work, no need for explanation (nicely written though). The ultimate question is whether there is any way to prove there is some kind of gain in slowing the memory down. With FSB at 413MHz I would be running the RAM at 826MHz to be synchro, which is way below its specs. And of course benchmarks report lower performance in such scenario (expected).

[jcarle]

I absolutely meant the first. Synchro with FSB.

Btw. You are talking about some prehistorical C2Ds The 65nm E6750 I got has default FSB of 333MHz.

I understand how this thing work, no need for explanation (nicely written though). The ultimate question is whether there is any way to prove there is some kind of gain in slowing the memory down. With FSB at 413MHz I would be running the RAM at 826MHz to be synchro, which is way below its specs. And of course benchmarks report lower performance in such scenario (expected).

Well, it's not just a question of prehistorical, it depends on the price point you bought at. Remember the E2160 (very popular, still sold a lot) runs with an FSB of 800MHz. It's a beautiful CPU to overclock with because you can run it at 1600FSB and get 1:1 with DDR2-800MHz memory.

Keep in mind, having an incredible amount of memory bandwidth is not very useful if your memory spends a great deal of time waiting for the CPU to come check it again. If anything, if your CPU is indeed the E6750, I would suggest overclocking the CPU to 1600(400)FSB which should be relatively easy, then running your memory at 800MHz. Using the new found overhead, it should be fairly easy for you to drop the timings down lower. Lower timings at 1:1 even with a slower clock will have a much greater impact on the performance then strictly playing with the clock speed.

[Octopuss]

Well, it's not just a question of prehistorical, it depends on the price point you bought at. Remember the E2160 (very popular, still sold a lot) runs with an FSB of 800MHz. It's a beautiful CPU to overclock with because you can run it at 1600FSB and get 1:1 with DDR2-800MHz memory.

Keep in mind, having an incredible amount of memory bandwidth is not very useful if your memory spends a great deal of time waiting for the CPU to come check it again. If anything, if your CPU is indeed the E6750, I would suggest overclocking the CPU to 1600(400)FSB which should be relatively easy, then running your memory at 800MHz. Using the new found overhead, it should be fairly easy for you to drop the timings down lower. Lower timings at 1:1 even with a slower clock will have a much greater impact on the performance then strictly playing with the clock speed.

But is there a way to measure it?

I currently run the memory at about 1030MHz. When I tested it and went to synchro ratio, it was already mentioned 826. I even managed to go down to CL3 and generally drop the timings by noticeablel amount, but benchmarks still reported pretty big drop in performance compared to the usual setup! So I really don't know what to think. I go by the numbers which happen to be uncompromising!

And um I already got effective FSB of 1652

So unless there is a test that can clearly show there is performance benefit in underclocking the memory, I guess I will stick to what I have at the moment.

Oh, let's have a screenshot of my usual configuration.

[Octopuss]

Look. Simple quick test:

[jcarle]

If you can run at higher FSB, then by all means go for it. All I'm saying is that when conditions allow it, you try to use the highest possible clocks at 1:1 for optimal performance, do keep in mind that it doesn't mean running at something other 1:1 is bad either.

[Octopuss]

If you can run at higher FSB, then by all means go for it. All I'm saying is that when conditions allow it, you try to use the highest possible clocks at 1:1 for optimal performance, do keep in mind that it doesn't mean running at something other 1:1 is bad either.

Yup I understand.

I am still dying to find someone to prove the main question about synchro mode giving better performance than higher asynchro speed

I gotta try to play around with getting stable FSB of 450+... That would really be juicy. I don't want to increase any voltages though - got passive cooling and don't need extra heat That's also why I run at pretty low frequency - 3.3GHz is max I can squeeze out of this cpu at default voltage.

[Octopuss]

Update!

It's true. I wouldn't believe it... Look:

And I'm not even at the highest possible frequency (didn't have time to play with it long enough).

[jcarle]

Yea, you managed to illustrate the point I was trying to make... 1:1 will outway the MHz benefit but only up to a point. You can't compare 1:1 with a 400MHz difference for example, but in a near range of frequencies, 1:1 will win.

[Octopuss]

Yup, I was pretty shocked seeing 1GB/s bandwidth increase for write. Who would have thought... Too bad I can't go higher on frequency as CPU wouldn't be stable afterwards Too bad the E8x00 series raised in price repeatedly around here recently.

Maybe I can try increasing voltage for the memory from 2.2V to 2.3 and see what timigs I could end up with