What is the formula for memory dividers?

[spittledip]

Hi,

i researched the web some but could not come up with a clear cut formula for working with memory dividers. What is the formula that determines the frequency RAM will run at when messing with the HT bus and CPU multiplier etc?

thanks

 

[n7]

How to calculate RAM dividers for A64 architechture:

(CPU multi x 200) / RAM divider (i.e. 166) = ___(always round this up if not whole number)

CPU multi x HTT / result of initial calculation = actual RAM speed

e.g.
OCed CPU clockspeed / result of the first calculation = actual RAM speed

one full example:

12 x 200 / 166.66 = 14.4 rounded up -> 15

12 x 220 / 15 = 176 aka DDR352
____________________________________________________


So take your case, we have 9x200 = 1800.
1800 / 400 (that's the divider you are using) = 4.5 rounded up = 5

To determine RAM speed, we then go...
1800/5 = 360 aka DDR2-720





This is just copied from one of my posts on another forum, since i'm not at home.

 

[spittledip]

Thanks for the help!

 

[MarcVenice]

http://math.gogar.com/athlon64.cgi?showtable=1

Just change the FSB speed accordinly and click on show ...

 

[DrMrLordX]

Keep in mind that, for some reason, Brisbane X2s run memory at different speeds than you might expect. My 3600+ ran memory at slightly slower speeds based on CPU multi (usually I got 95% of the expected speed when using the default 9.5x multi).

 

[FireChicken]

what about for C2D My friend is running at 3.44 using 1:1 dividers on a DS3 usinf ddr2-800. What what would the ramspeed be? thanx

 

[n7]

On Intel, dividers are easy.
(Though you can run asyncronous dividers on nF chipsets.)

1:1 just means the FSB & RAM run at the same speed.
4:5 means RAM is 1.25x faster than the FSB.
2:3 means 1.5x faster, & so on.

Originally posted by: FireChicken
what about for C2D My friend is running at 3.44 using 1:1 dividers on a DS3 usinf ddr2-800. What what would the ramspeed be? thanx

Depends what CPU & multi he's running.

 

[MarcVenice]

err firechicken, it totaly depends on the FSB of your cpu. If I assume its a e6400, with a 8 multie, and his FSB is 430, which would give us 3440mhz, then his ramspeed would be 430*2 = 860.

1:1 simply means your ramspeed equals the FSB times twice. Its doublepumped since it's ddr2, which I can't really explain myself, I just know its true

 

[n7]

For those of us who have a hard time translating what we know simply: http://en.wikipedia.org/wiki/DDR2_SDRAM

 

[FireChicken]

Thanks for the info guys and yes he is running an e6420 430 FSB with an 8x multi

 

[myocardia]

Originally posted by: DrMrLordX
Keep in mind that, for some reason, Brisbane X2s run memory at different speeds than you might expect. My 3600+ ran memory at slightly slower speeds based on CPU multi (usually I got 95% of the expected speed when using the default 9.5x multi).

No, they don't work any differently than all other A64's. But, all A64's do that when using a ½ cpu multiplier.

 

[DrMrLordX]

That's not true; the Brisbanes are different. The link you posted applies to s939 and s754 K8s using the older DDR memory controller. I observed the following on my Brisbane chip:

when using the 1:1 ratio, there is no difference in memory speed based on multiplier, nor is there ever any penalty even when using half multipliers. Example:

HTT 200 x 9.5, 1:1 = DDR2-400
HTT 200 x 9, 1:1 = DDR2-400
HTT 200 x 8.5, 1:1 = DDR2-400
HTT 200 x 8, 1:1 = DDR2-400

When using any ratio other than 1:1, the speed varies based on multiplier, though not in the way you'd think. The closest way I can represent the actual memory speed is as follows:

((CPU speed / CPU divider) * ratio) * (CPU divider / 10)

So I'd get stuff like this:

HTT 200 x 9.5, 1:2 = DDR2-760
HTT 200 x 9, 1:2 = DDR2-720

Windsors had their own quirks which were documented in some of the original AM2 articles on Anandtech (too lazy to look them up since Windsors aren't the chips in question here).

 

[myocardia]

Originally posted by: DrMrLordX
That's not true; the Brisbanes are different. The link you posted applies to s939 and s754 K8s using the older DDR memory controller. I observed the following on my Brisbane chip:

When using any ratio other than 1:1, the speed varies based on multiplier, though not in the way you'd think. The closest way I can represent the actual memory speed is as follows:

((CPU speed / CPU divider) * ratio) * (CPU divider / 10)

So I'd get stuff like this:

HTT 200 x 9.5, 1:2 = DDR2-760
HTT 200 x 9, 1:2 = DDR2-720

Hmm, it's crazy, but you're right. I hadn't even checked into that, since I wasn't the slightest bit interested in a Brisbane, and already knew how AMD's memory subsystem (used to) work.

 

[DrMrLordX]

It does seem crazy, but my guess is that it's a part of a power-saving system.

Cool&Quiet downclocks the CPU to save power, but your memory keeps trucking right along at the same speed, right? Well, I theorize that Brisbanes have a multiplier-based speed penalty on RAM to allow Cool&Quiet to downclock your CPU and memory at the same time. How much power that really saves, I don't know, but it would probably save you some.

It also makes overclocking these little chips a bit weird, but it's not a major impediment.

edit: I'd also like to point out that my formula may only be applicable to the 3600+ with its sub-10x multiplier. I have no idea how it works on Brisbanes with multipliers higher than 10x.

 

[myocardia]

Originally posted by: DrMrLordX
edit: I'd also like to point out that my formula may only be applicable to the 3600+ with its sub-10x multiplier. I have no idea how it works on Brisbanes with multipliers higher than 10x.

Then you should read the link I supplied (the very bottom of the page).

 

[DrMrLordX]

Originally posted by: myocardia

Then you should read the link I supplied (the very bottom of the page).

I did, but the results they posted didn't seem to follow any logical order that I could immediately discern (read: I'm lazy and wanted them to match the data to a formula). It would have been nice had they posted results based on ratio + multiplier rather than CPU speed, not that it was difficult to figure out which multiplier they were using.

It seems almost like they're suffering from the old Windsor odd multiplier bug plus another speed hit for running a half-multi, but without more data I can't say for sure. Interestingly enough, I did not have the odd multiplier bug on my 3600+ when using an 8.5x or 9x multi. Weird.