Complete A64 Memory Divider Table

[OSKAR_WU]

Complete A64 Memory Divider Table

Ps : It may not be 100% correct

http://oskarwu.myweb.hinet.net/A64Divider.jpg

[boshi]

umm... ok...

cool

:confused:

[Karnivore]

Thanks Oskar, nice bit of info :thumbsup:

[OSKAR_WU]

Originally posted by OSKAR_WU
Complete A64 Memory Divider Table

Ps : It may not be 100% correct

http://oskarwu.myweb.hinet.net/A64Divider.jpg

Any problem from this table ? :confused:

[texuspete00]

Nice to see you again OSKAR. This will come in useful for my yuletide A64 system.

:toast:

[CodeRed]

Better off presenting a table of memory dividers.

125 MHz is actually 117 MHz AFAIK
No idea about the 143 MHz setting.
in goes in steps of 1/12 ;)

[OSKAR_WU]

Originally posted by CodeRed
Better off presenting a table of memory dividers.

125 MHz is actually 117 MHz AFAIK
No idea about the 143 MHz setting.
in goes in steps of 1/12 ;)

I think you know DRAM frequency we set in A64 register is " Upper limit " of DRAM speed ...

A64 memory controller will automatically choose a smallest interger divider(>4) to generate the DRAM clock from Core Frequency , and the DRAM clock will not be faster than the uppder limit ...

Using an oscilloscope and check all the multiplier from 4~12 with the same register setting in A64 ...

What I got are 200/183/166/150/143/125/100 may be the upper limit of DRAM speed when you set correspoding register setting in A64 ...

Here is the divider table I got after using an oscilloscope to check the possible upper limit of DRAM clock ...

What I can not be sure is the exact frequency besides 200/166.67/133.34/100 ...

http://oskarwu.myweb.hinet.net/A64D1.jpg

[tictac]

wow.... :)

[tritium]

That looks nice, but since I do not have an A64 yet can you please explain the chart. Memory speed, HTT setting , CPU ratio etc....

Thanks :) :)

[Dissolved]

Tagged..

[Edward Ng]

Sorry, but I need to ask the obvious question...

Any way you can find out or tell us what happens to the RAM Divider for DRAM Freq. over 200 in BIOS? For example, if I use DRAM Freq. 225, 250, 275 or 300?

Thank you for any information you can provide, and furthermore, thank you so much for your efforts, Oskar. Your LP nF3 is a true marvel of modern engineering!

-Ed

[craig588]

You can't use a devidor less than your multiplier.

[Edward Ng]

Okay, suddenly I understand...

I was thinking of the LDT setting, not the RAM ratio.

Gotcha'; thanx.

[saaya]

wow.... :)

ditto!

thx a lot oskar! :toast:

[craig588]

Yeah, I love this divider table.

[Çhrist0ph]

where did the divider table go? Can someone post it again?

[Learn]

where did the divider table go? Can someone post it again?

*.png8 139KB & 155KB :D

[papatsonis]

it simpler if u don want to remember tables.

u multiply the (HTT/ram) ratio -inverse of ram:htt ratio in the bioses- with the
cpu multiplier , and u aproximate to its closer bigger integer.that's ur memory divider,then u divide ur cpu mhz by this divider..

example if multi is 8 ,HTT=240(=>cpu=1920MHZ) and divider "ram:htt" is 5:6 , u multiply 8*(6/5)=9.6 , then approximate to the bigger closer integer =.. 10 and divide the cpu clock by ths number ...=> ram clock is 1920/10=192mhz

of course if u choose a half multi , the bios "approximates" it to the closest integer(bigger)

[FrozenGPU]

it simpler if u don want to remember tables.

u multiply the (HTT/ram) ratio -inverse of ram:htt ratio in the bioses- with the
cpu multiplier , and u aproximate to its closer bigger integer.that's ur memory divider,then u divide ur cpu mhz by this divider..

example if multi is 8 ,HTT=240(=>cpu=1920MHZ) and divider "ram:htt" is 5:6 , u multiply 8*(6/5)=9.6 , then approximate to the bigger closer integer =.. 10 and divide the cpu clock by ths number ...=> ram clock is 1920/10=192mhz

of course if u choose a half multi , the bios "approximates" it to the closest integer(bigger)

Wow, that IS a lot simpler. :confused: :stick:

[FoxTrottZero]

I'm new here and to OCing... but I have a theory about pushing the Hz and Voltage on the system:
I think that these tables indicate it off the bat, but what I imagine is that it really helps to keep the frequencies in as efficient a synch rate as possible...
Looks like... a 4Ghz system would be most stable if one were to use RAM of as compatable a frequency... something like 400Mhz
That way, the clock pulses have a pretty good chance of interlacing properly.

From what I understand, AMD's FSB controls the RAM directly from the CPU, so it would make sense then to make sure that you can run compatible RAM. Seems like the system would be better of say at 4Ghz then at 4.1 with 400mhz ram in terms of stability.
Now Raw speed... that's another thing ;)
Just an idea, I really have no expertise on this.

BS?

[p4z1f1st]

oskar, or somebody else:

could you tell me the "interval" with which you calculated that all?

because, i want to extend the table with clocks like 217/233/250/266/275/283/300MHz ;)

[p4z1f1st]

ok, got it:

multiplier * your wanted FSB(HTT)-clock / wished divider (f.e. - 166 for the 6:5-divider or 133 for the 3:2-divider and so on) = a real-count

you have to "up-round" the result - so, if it's 13,2 you have to say, it's 14 - if 12,6 it's 13, if 12,1 it's 13, too

so, then:

CPU-Clock / result of the first calculation = REAL RAM-CLOCK


one full example:

13 * 200 / 166 = 15,66 ---> = 16

2600 / 16 = 162,5MHz

http://www.8ung.at/dresign/a64divider.jpg

[saaya]

isnt it funny that it goes to 5ghz cpu speed but only 200mhz memory speed? ^^

[papatsonis]

i have some ''objections'' tο the above table

it needs to be clear enough, what exact every ratio is,e.g. 166/200=5/6 , but for the not common ones (183,143) , is more ..misty

the 183/200 what should be? 183.3333333/200 which means 11/12 ??? ,should be 9/10 as dfi says in the bios??? , or 10/11 ???
From the table , we can find some multis that are in the ''edge'' to change for different ratios (9/10, 10/11, 11/12) so the rounding could be different ,and IF , we assume the table is correct we might find the right one.

--from now on will be LONg post :)

for 10x multi(CPU) 10/9 gives 11,1111-> 12x ram multi
11/10 gives 11 -> 11x
12/11 gives 10.9-> 11x

for 19x , 10/9 =21.1111-> 22x
11/10=20.9 -> 21x
12/11=20.73->21x

for 20x , 10/9=22.2 ->23x
11/10=22->22x
12/11=21.81->22x

for 21x , 10/9 =23.3->24x
11/10=23.1->24x
12/11=22.9->23x

for 22x , 10/9=24.4->25x
11/10=24.2->25x
12/11=24->24x

so.. from the table and the calculated ram mlutis (for cpu 10x,19x,20x) we can easily see, that the 10/9 is unacceptable, so 11/10 or 12/11

but...for 21x multi =4200MHZ clock , for 183mhz ram setting ,the table says 182.61MHZ =..23x ram multi as for 11/12 ratio

and for 22x =4400mhz >> , >> 176mhz, as for 11/10 ratio..!!?? whats happens here??of course IMO , the correct ratio is 11/12 as for 183.33333333/200

doing the same thoughts in propotion for the 143mhz divider , i think -and seems to be confirmed the 5/7 ratio ,

2 more observations.. :

1)all the mem dividers for half multi dividers, are ''calculated"-oscilloscoped from the above table, like the half multis do exist, but... we all think that they are just a trick to give more flexibility.example from the table : 6.5x cpu multi , 5/6 ratio (should be =rounded to 7*1.2=8.4=..9x ram multi) ,but it gives 6.5x1.2=7.8=..8x ram ratio.Also note that for 1:1 setting the half multis are normally up-rounded .. so may that mean that the half multis DO EXIST when not using 1:1 htt:ram setting???
2) the cpu multies <5x seems to be emulated using 5x multi , as a consequense the 4x multi 1:1 gives only 160mhz!!

[tictac]

0.5multi bug.... :(

Oskar wu can you fix it? :)

[jlccarv]

Sorry Oskar...but I don't think this chart is accurate. Why didn't I gain bandwidth at the same HTT speed by changing to a "solid" multi ???

Lets bust up another myth or get to the bottom of it...

-Holly

http://xtremesystems.org/forums/showthread.php?p=613669&posted=1#post613669

RE: Your linked post...

CPU is 9.5*280=2660, with a 9.5 CPU multi the Memory is 2660/10 = 266

CPU is 9.0*280=2520, with a 9.0 CPU multi the Memory is 2520/9= 280

You have 140MHz extra clock on the CPU with the 9.5 CPU multi. That would give you the increase in the Buffered bandwith. Remember that CPU clock affects buffered bandwith...

[Badge56]

Am I reading this correctly? The RAM will not go over 200 mhz no matter what divider you use on the CPU?

I am at 2450 (10X245) RAM is DDR33 so its at 204 mhz
at 2400 (10X240) RAM at DDR333 it was also at 204

Will not post if RAM at DDR400

Explain please.... this is NOT clear yet.

[jlccarv]

Also...If it's true, I can run Prime all day long at 2.5,3,7,3 at 266HTT. I got instant Prime AND SuperPi errors the minute I tried 2.5,3,7,3 at HTT280 with BOTH 9x and 9.5x multis.

Have you considered that its your high HTT that maybe the cause for your errors in prime and superpi at 280HTT. Or even your HTT multiplier maybe too high...

[hollywood]

Am I reading this correctly? The RAM will not go over 200 mhz no matter what divider you use on the CPU?

I am at 2450 (10X245) RAM is DDR33 so its at 204 mhz
at 2400 (10X240) RAM at DDR333 it was also at 204

Will not post if RAM at DDR400

Explain please.... this is NOT clear yet.


the 200 means 1:1 ratio to HTT speed...

[CodeRed]

It may, but not unbuffered. I saw no improvement going to ANY solid multi compared to an x.5 one.

Also...If it's true, I can run Prime all day long at 2.5,3,7,3 at 266HTT. I got instant Prime AND SuperPi errors the minute I tried 2.5,3,7,3 at HTT280 with BOTH 9x and 9.5x multis.


Again...I think that the table is either outdated, mis-interpreted or simply BS...I simply have found zero evidence that I'm losing either HTT speed or bandwidth within a variety of x.5 multi combinations compared to their "solid" brethren.

You must be the only one that doesnt believe in the 0.5x multi thing. Here are some pifast tests that show the multi is being rounded up:

2690 MHz (10x269) 43.38 sec
2689 MHz (10.5x259) 43.89 sec
2692 MHz (11x244) 43.89 sec
2689 MHz (11.5x233) 44.38 sec
2691 MHz (12x224) 44.39 sec

As you can see the 10.5 result is the same as the 11x result and so too with 11.5x == 12x


There is no doubt that the mem controller rounds up the 0.5x multipliers.

[hollywood]

You must be the only one that doesnt believe in the 0.5x multi thing. Here are some pifast tests that show the multi is being rounded up:

2690 MHz (10x269) 43.38 sec
2689 MHz (10.5x259) 43.89 sec
2692 MHz (11x244) 43.89 sec
2689 MHz (11.5x233) 44.38 sec
2691 MHz (12x224) 44.39 sec

As you can see the 10.5 result is the same as the 11x result and so too with 11.5x == 12x


There is no doubt that the mem controller rounds up the 0.5x multipliers.

So now it rounds up ??? Everyone is saying that it rounds down...even that chart says it rounds down.

pfft...I'm getting a migrane... :(

[CodeRed]

So now it rounds up ??? Everyone is saying that it rounds down...even that chart says it rounds down.

pfft...I'm getting a migrane... :(

No the chart shows that it rounds up.

If it rounded the memory multi down then the mem speed would exceed the spec values of 200, 183, 166 etc etc. Thats why AMD had to implement it this way.

[hollywood]

Yeah...I get it...

CPU at 9.5 multi x 280HTT = 2660

Mem multi goes to 10

divide CPU speed by 10 = 266

Means "real" HTT is only 266.

I was wrong and stand corrected, thank you for helping me understand CodeRed and jlccarv.

So now the question is...how high can I get at only 9x, cause 10x divider put my mem too low... :(

[Badge56]

Ok, let me get this straight.
My current set up is 2450mhz (245X10) DDR333
System shows 2450mhz CPU and 204 mhz RAM

According to the chart, using DDR333 (aka 166 mhz) the divider should be 12 and it is according to CPUz but my RAM is running 204mhz ????? According to the chart is should run 160 mhz

What am I not understanding?

[papatsonis]

no , its correct at 204, 10x *(200/166)=12 = your ram divider, 2450/12 =204,....

[Badge56]

Yep that part is clear. Why does the chart show Real dram frequency = 166,67 for 2500? Why not 208 ( 2500/12 ) Why the 12.5 divider?
According to CPUz I am now at 2500 mhz and RAN 208

[bias_hjorth]

Hmm I am having a hard to figuring out what actual speed my modules are running. Hope someone cloud clerify this.

11 x 274 (with the 183divider) - Buggy eh? :hehe: Or is it just me :(

[CodeRed]

Hmm I am having a hard to figuring out what actual speed my modules are running. Hope someone cloud clerify this.

11 x 274 (with the 183divider) - Buggy eh? :hehe: Or is it just me :(

mem multiplier is 200*11/183.333 = 12 exactly

so mem speed is 11x274/12 = 251.67 MHz

[bias_hjorth]

:thumbsup: thx

Although I may have hit some kind of bug cause my memory cant run at that speed. They usually max out at 245 and is very unstable up from 240.

*edit if I bootup at 166mhz mem speed the bios shows me the correct 227mhz at the 274 x 11. But cpu´z and everrest shows me 217mhz.. Seems fishy :confused:

* edit 2: Well after seeing the no 1 table again I think you may have miscalculated.
11 x 200 = 183 divider choosen = ~ 169mhz real speed
------------

11 x 200 = 2200 / 169 = ~ 13
11 x 274 = 3014 / 13 = ~ 231

Or can I be wrong here. Btw. notice according to the no 1. chart the 11.5 multi is faster than 11 and 12 :)

[papatsonis]

yes, eventhough till yesterday, i was thinking that the ratio was 11/12 , now i think its 10/11 , cause a person i know was bnchmarking at 11x297 with 183 divider, and no way (even he, admitted that) he could have his BH-5 at 3266/12=271.... more probably its 3266/13 , so its 11*(11/10)=12,1 rounded o 13

[CodeRed]

[QUOTE=bias_hjorth* edit 2: Well after seeing the no 1 table again I think you may have miscalculated.
11 x 200 = 183 divider choosen = ~ 169mhz real speed
------------

11 x 200 = 2200 / 169 = ~ 13
11 x 274 = 3014 / 13 = ~ 231

Or can I be wrong here. Btw. notice according to the no 1. chart the 11.5 multi is faster than 11 and 12 :)[/QUOTE]

I didnt use the table I used the formula:

mem multi = ceil(200 * CPU multi / mem speed ratio)

Maybe it is 13x instead of 12x since its on the edge. Might be a good idea to test with something like pifast

try 13x200 1:1 vs 12x217 1:1 vs 11x236 @ 183 and see which are equal.

[Formann]

Does this make it any easier?

Im pretty sure this is correct.

http://xtremesystems.org/forums/showthread.php?p=628459#post628459

[Capt MonkeyBoy]

hi'ya



i made an App using

13 * 200 / 166 = 15,66 ---> = 16

2600 / 16 = 162,5MHz

and it fits the table's on the first page of this thread, with all it's calculations at 200mhz, but it differs from Z-Cpu and the bios at higher HTT settings..??

the way the calculation is worked out it will hold the Ram at the frequ you set the divider at.....

like this

13 * 250 / 166 = 19.57 ---> = 20

3250 / 20 = 162,5MHz

is this right....or will it float up with the HTT,,,,,



cheers

[AKRedneck]

Or try the spreadsheet cause it can't get any easier...complete with graphs, etc in the separate sheets. Plug yer numbers and it does the rest..

http://www.xtremesystems.org/forums/showthread.php?t=49393
:D

[boblemagnifique]

hello

could I know has how much is my RAM really please ?

thank :toast:

http://www.membres.lycos.fr/boblemagnifique/Overclocking/A64/cbbhd0447tpew/superpi35s5.PNG
http://www.membres.lycos.fr/boblemagnifique/Overclocking/A64/cbbhd0447tpew/superpi31s2.PNG
http://www.membres.lycos.fr/boblemagnifique/Overclocking/A64/cbbhd0447tpew/superpi29s2.PNG

[SpotTheCat]

http://img.photobucket.com/albums/v344/Sp0tTheCat/overclock.jpg
as you can see, I have a really nice overclocker in my hands. The problem is I can't for the life of me get my memory to it's rated speed. The memory dividers in my bios are only 133, 166, and 200 (14, 11, 9)
my memory (corsair value PC3200 2.5cas) can only make it to about 205mhz, so right now I am using the 14 divider that gives me 186mhz memory, could I change it to 13 so I can have ~200mhz memory again?

[brandinb]

lol for you guys having problems figuring this out go to oskars second table and find the multi you are using. then follow it to the right and go to the ration your using. then devide your total cpu mhz by that number for your final ram speed.
________
FeedMyLips (http://www.girlcamfriend.com/cam/FeedMyLips/)

[bias_hjorth]

Yeah sure brandinb.. Thats the way to do it..
This is the correct way:


Ex
11multi x 200htt = 2200overall mhz / 169 from oscar_wu = ~ 13 mem divider.
11 x 274 = 3014 / 13 = ~ 231

[SpotTheCat]

I know how to calculate everything, but I don't know how to get a memory divider of 13 on my GA-K8NF-9
only options I can find are 133, 166, 200

[Unhooked]

i think CPuz too gives you the correct readings for ram
here is mine
http://ocmods.com/280fsb.jpg

[Orthogonal]

I know how to calculate everything, but I don't know how to get a memory divider of 13 on my GA-K8NF-9
only options I can find are 133, 166, 200

Unfortunately, not all motherboards give all the different multiplier options.

[DarkManX_BG]

CPU-Z reads it right for me:
I run 300HTT, a multi of 8x and I want 250MHz memory speed. The 166 (5/6) divider should give me 250MHz. However:

200x8 / 166 = 9,xx -> 10
2400/10 = 240MHz memory speed
And that's what CPU-Z reads here...

[P_1]

Or try the spreadsheet cause it can't get any easier...complete with graphs, etc in the separate sheets. Plug yer numbers and it does the rest..

http://www.xtremesystems.org/forums/showthread.php?t=49393
:D

So the neo2 wont work with a 9 multi and the 133 divider? (im trying to get 2800mhz on cpu with ram running at 200 on a 3000+ thats limited to a 9 multi on a neo2)

[Spajky]

How to know the real memory clock on AMD "K8" platform:

... { A64´s have a default clock reference [System Bus] at 2HGz & default mem. divider @ 10, while s.754 Sempron´s at 1,6GHz & divider of 8. That means for Example : running a 50% OC-ed 1,8GHz CPU @ 2,7GHz-(Htt-300MHz) : Sempron with mem.divider 9 & A64 with 11; real memory clock on first one is 267MHz, while on another is 273MHz, if your memory sticks can handle it } ! ...

So; for A64: Htt x [def.div(10) / mem.div.] = real mem.clock
& Sempron: Htt x [def.div.(8) / mem.div.] = real mem.clock

CPU-Z is not reliable on AMD "K8" platform showing memory clock (mem.divider is shown properly!)


My (research) article "about BENCH´ING MEMORY -real life"
with WinRAR´s built_in benchmark & hardware test" - on my site: :)
some tests/benchmarks & explanation HOW IT WORKS, is here directly:

http://freeweb.siol.net/jerman55/HP/benchMem.htm

[especially the yellow column is interesting!]
/sorry, I didn´t make any graphs/

Regards ...

[T_M]

*breaks out the shovel for some digging*....

Only thing i can find wrong with the table Oskar is that you spelt "Ratio" wrong ;)
Apart from that is seems spot on for my own findings on 10x and 11x.

[ImportantAwareness89]

Ok, so Athlons use this type of divider setting, but Intels still use the simple FSB * ratio right? Only Athlons involve the whole ceiling, multiplier, etc...??

[lukija]

Hi. My first post here :)

I wrote small utility , you can grab it here:

http://users.yubc.net/~lukija/A64Info.rar

It is still unfinished, untested .... works well on my Winchester. Use it on your own risk. Yes, i know, it has stupid name (A64Info) :) You can tweak some values like in A64Tweaker and it also has built in calculator which calculates memory speed and I used Oskars table as a starting point.

But, I have one question regarding undocumented dividers. For example 180 ( from the DFI point of view). I'm not sure if it is 183.0 MHz or 183.33..... MHz. According to the table, it is 183.0 MHz. For example, you have 2200 MHz CPU (200 MHz x 11). And use 180 divider on DFI. If we look at the table or use calculator in A64Info we get divider 13 and memory speed of 169.23 MHz. But if we assume that 180 divider implies 183.33..... MHz then we get 12 divider and exactly 183.33..... MHz. It seems logical to me. It would be interesting to measure memory bandwidths in these two scenarios:

1) CPU Speed 2200MHz (220 MHz x 10). Using 166 divider we have 12 memory divider which finally gives 183.33 MHz memory speed. Measure bandwidth using Sandra or Everest. This bandwidth will be used as reference.

2) CPU Speed 2200MHz (200 MHz x 11). Using 180 divider we have:

a - 13 divider (if we assume thah upper limit for memory speed is 183.0 MHz) and memory speed of 169.23 MHz

b - 12 divider (if we assume that upper limit for memory speed is 183.33... MHz) and memory speed of 183.33... MHz

Comparing measured bandwidth in case 2 with one measured in case 1, we could see what is upper limit. If bandwidth in case 2 is equal (+/- few MB/s) to one in case 1 we can say that upper limit is 183.33... Mhz. If bandwidth in case 2 is lower ( because of 14 MHz lower memory speed) then one in case 1 we can say that upper limit is 183.0 MHz.

Required: CPU with at least 11 multi, mobo with 180 divider, Sandra, Everest.

Any ideas, suggestions, comments?

[ReelMonza]

Nice, will check your app later, can you please look at this (http://xtremesystems.org/forums/showpost.php?p=1046541&postcount=390)

[lukija]

Well friend of mine did some tests, CPU is A64 3500+ Newcastle core:

a) 220x10 , 166 divider (183.33 MHz memory speed), using it as a reference value since 166 divider is "clean"

Sandra - 5518/5480 MB/s (integer/float, buffered,SSE2)
Sciencemark - 5079.09 MB/s

b) 200x11, 180 divider, (memory speed is either 183.33 MHz or 169.23 MHz)

Sandra - 5122/5114 MB/s (integer/float, buffered,SSE2)
Sciencemark - 4750.25 MB/s

Since bandwidth is noticeable lower, I think that memory is running at 169.23 MHz thus upper limit for memory clock for 180 divider iz 183.0 MHz rather than 183.33 MHz.

Still I'm waiting for one test - 220x10 with 150 divider which will give memory speed of 157.14 MHz, just to have another reference value to comare with.

Any suggestions?