AMD CPU o/c analysis

**MrSeanKon** · 09-21-2006, 10:25 PM

1) The purpose of thread
This thread is for all AMD users.
I will show the overclocking behavior of many AMD CPUs.
Although I share Guidemania (many o/c guides inside) it is impossible to add in this program anything --> LOL enormous size RULEZ --> people will not download it!
On the other hand it is necessary sometimes to discuss with users the results.
Hopefully after some years (long database is required) hope to present here general results.
This thread is for practise also.
I add the necessary photos and explain what is going on so your experience will be higher during reading the thread.

2) Prerequisites
You should download the Overclocking Bible (OcBible shortly) at least.
Guidemania is also recommended especially if you are not familiar with many ideas.
Start reading "The basics" (Part 1 of Top o/c guide) and then Part 9 (Overclocking analysis) just to be familiar with the following stuff.
Don't be afraid it is not necessary to have a Ph.D. degree in Mathematics to understand some texts.
I am not Mathematician!
--> Note that OcBible includes my personal AMD database where I present some details for all processors.
--> Download the attached file (IT IS USEFUL ONLY for OcBible users).
It is my personal AMD database (O/c analyzer exponent files) thus you save time of typing my results.

3) Links for the programs
You can download the latest (and older versions) of Guidemania and OcBible at Aoaforums.
Both programs are freeware of course.
Click here/hier clicken to grab all editions are downloadable for any user.
Enjoy!!!!

4) Supported threads
Both programs are supported in many other forums; ABIT-USA, ABXzone, Aoaforums (the homepage), OCAU, Ocforums, PcPer, XtremeSystems (obviously) and Greek users can post at Insomnia forum.
Where are the links??????
See my profile here --> for the XtremeSystems OcBible supported thread.

5) Acknowledgements
Thanks to:
1) All people who download both programs and post kind comments.
2) The OCAU users for OCAU sticky there, yeah folks you have posted many links!!!
3) The Aoaforums for hosting both programs.
4) My cooperators _dangtx_ (PcPer), giorgos th. (Insomnia and TechPowerUp!), Mad 4 Air (OCAU), SoF (GKR I hate you LOL) and zeppelinrox from AMD forums.
If someone is not listed let me know or u2u me.
Don't you know what does u2u mean???
Download Guidemania --> Reference guide --> Part 1!

Ladies and gentlemen:
Fasten your belts!
We will fly to the brilliant AMD o/c world!!!!!

**MrSeanKon** · 09-21-2006, 10:26 PM

1) Athlon 2200+
This processor tested using different heatsinks.
Thus the temperatures are totally different --> look my personal AMD database for more info.
First measurements:

Second:

Comparison:

As you see the distance between two curves is almost the same.
This means that temperature did not play any role for final MSFs.

Questions:
1) Why Logarithmic fit???
The reason is based on analysis of curves and some other observations.
It is beyond of my scope to present here this stuff.
If in the future change something in internals I will inform you.
2) Where OcBible uses Logarithmic fit?
Run Data Modeling and perform Logarithmic fit for some MSFs and Vcores (hypothetical values).
Then run AMD CPU speed and choose the Regression combobox.
Same results of course!

**MrSeanKon** · 09-21-2006, 10:29 PM

1) Barton 2500+ ABIT mainboard
Like above --> different temperatures and Vcores.
First measurements:

Second:

Comparison:

The difference between two curves is the same (two white arrows show it) but after second white arrow the navajo white curve overtakes the red.
The yellow arrow shows the third pair where Vcore was 1.962 and after this point LOL banana convergence!

Exercise:
Perform 2nd or 3rd degree polynomial fitting for first set --> open Data Modeling form.
You must set a different value for Vertical shift e.g. 2 for example just to see the whole curve.
The shape seems banana or not????

Special notes:
1) Focus in first photo (the second partial exponent).
Small value don't you think????
I BANNED this pair from OcBible internals.
You must realize that program cannot predict well for pathological cases like that.
2) Compare the total exponent n6 (2284 --> 2520 MHz) and n7 (2284 --> 2541).
They are similar!
Remember what is BAN method --> read in Guidemania --> Part 10 of Top o/c guide.
Why pn6 and pn7 (pn means partial exponent) are totally different???
The reason is that any pn is defined using logarithmic differences between current MSF and previous.
Thus a small rip on curve (or on derivative this is a better approach) leads to high pn difference.
On the other hand n (the total exponent) is defined between current MSf and the first one.

**MrSeanKon** · 09-24-2006, 10:29 PM

3) Barton 2500+ Epox mainboard
Some measurements were taken using Vcore < of Barton default value = 1.65 Volts.
CPU worked just below of Breakdown Region cos 1.857 / 1.65 = 12.54% and full load temperatures were < 45 Celsius.

Exponents:

And the total curve:

The yellow arrow shows the point (saddle point) where the % loss changes sign.

Let's calculate the exponents starting near to default Vcore = 1.65.
MSFS and Vcores:

Although the n4 > n3 (0.603 > 0.571) % loss is positive but lower (18.592 vs 23.253).

Below this table used for variation analysis.

All exponents are > 1 for Vcores <= 1.574.
This is normal although was aircooled note that subzero cooling improves the exponents but I do not have one.
Hopefully in the future!
If you perform fit for five first points --> you see the total curve has different carvature.
For Vcores < 1.46 has exponential shape and above 1.46 Volts logarithmic.
Thus the saddle point is somewhere to 1.46 Volts OK I do not care for the exact value cos any model gives different values!
When n flutters all CPU estimators are not accurate.
More fluctuation --> less predictable CPU --> do not use BAN method.
For this reason when you start finding the first MSF for aircooled and watercooled systems set Vcore closed to default value.

**Lightman** · 09-25-2006, 01:27 AM

Good work! Can you do same measurements for K8 CPUs??
I'm wondering how this architecture on 90nm will behave compared to K7 on 130nm.

Thanks!

**aspevacek** · 09-25-2006, 06:57 AM

Why not compare the K8 on 130mm to make the test a little more even. Then you can work with the different revisions and Die Shrinks. It would make it alot easier to see where the gains were made

**MrSeanKon** · 09-28-2006, 12:54 AM

Thanks for comments!
My current CPU is San Diego 3700+ so you should not be impatient wait until my next upgrade!
When I buy a new one I will post info for this.
Now I am posting the necessary info for old fashion AMDs.

4) Barton 2600+
Similarly with first sample (Athlon 2200+) this processor tested using different heatsinks.
First measurements:

Second:

Comparison:

The dark orange arrows show two points which Vcores were similar (1.732, 2295) vs (1.733, 2354) and (2.078, 2435) vs (2.086, 2526).
The first comparison shows 2354 - 2295 = 59 MHz difference but the second 2526 - 2435 = 91.

FIRST RULE: If CPU temperature does not effect to MSFs then their difference is constant regardless Vcore and cooling.
As you understand this is not true always.

**MrSeanKon** · 10-09-2006, 11:05 PM

5) Barton 2800+
Measurements:

Exponents:

Total plot:

Logarithmic shape Banana convergence!

**MrSeanKon** · 10-09-2006, 11:07 PM

General results for Bartons and XPs
Let's perform some statistics using all o/c exponents (only inside Normal Region) --> see attachment photo!!

WOW!!!!!!!!!!!!!!!!
The possible value for o/c exponent is 0.603 --> almost 0.6!
Believe me I did not know this!!!!!!
Maximum value is .816 and minimum .403 but many values are around 0.6.
Yeah:
1) In [0.55, .65] region --> 10 samples of total 23 --> 10/23 = 43.4% possibility.
2) In [0.5, 0.7] region --> 17/23 = 73.9% possibility.
Thus for aircooled/watercooled XPs and Bartons (desktop CPUs only) 0.5 <= n <= 0.7 in most cases.

For Barton 2500+ (AQYFA stepping - Epox mainboard) the first exponent was outside of region [0.5, 0.7] as you saw.
It was 0.744 (2320 --> 2365 MHz).
The next exponent was 0.595 --> difference = 0.744 - 0.595 --> 0.149 but Vcores must not change rapidly OK?? --> Mapping method.
For AQXEA Barton 2500+ exponent in the beggining was 0.816 and became (the second value) 0.682.
Difference is 0.816 - 0.682 = 0.134 which is near to 0.149!!!!!!!!!
Why not guess 0.14 for difference????
Therefore if XP/Barton o/c exponent is higher than upper limit = 0.7 then the next exponent will be 0.14 units lower only for aircooled/watercooled desktop CPUs.

**MrSeanKon** · 10-09-2006, 11:08 PM

General results for A64s (754 and 939 single core)

Well it is not necessary to upload many photos again I think you understand the process!
In addition you can download my personal AMD database for practise.
Therefore:
754 3200+ Albatron (first case)
Exponents are 0.513, 0.502, 0.48, 0.43 and 0.382.
754 3200+ Albatron (second case)
--> 0.681, 0.493, 0.52, 0.509, 0.387, 0.317 and 0.305.
754 3200+ Epox
--> 0.925, 0.689, 0.672, 0.62, 0.502, 0.471 and 0.469.
939 3200+ Winchester
--> 0.912, 0.761, 0.764, 0.749, 0.637 and 0.593.
939 3200+ Venice
--> 0.908, 0.927, 0.838, 0.831, 0.772 and 0.685.
939 3700+ SanDiego
--> 0.519 and 0.523
939 4000+ SanDiego
--> 0.623, 0.482, 0.496, 0.483 and 0.437.

Run variation analysis again entering the above values.
You will see:
O/c exponent for 754s possible value = 0.52

O/c exponent for 939s possible value = 0.68

O/c exponent for all single core AMDs possible value = 0.6

WOW!!!!!!!!!!!!!!!!!!!
Like XP + Bartons!!!!!

AMD GOLDEN RULE FOR ALL SINGLE CORE CPUs
If you have a desktop AMD single core which is aircooled/watercooled which is not working inside Breakdown Region then the exponent n is around 0.6.
The possibillity for [0.5, .7] area is 49.2% or 30 samples of total 61.
Hope it helps

More results and comparisons in the future when I collect more data of new generation CPUs.

**SoF** · 10-10-2006, 01:23 AM

Why don't you add more coefficients in general approximation models?
Just to improve the convergence and CCs.
In most cases the specific AMD estimators predict better than General CPU speed estimator.

**MrSeanKon** · 10-11-2006, 10:52 PM

Good question SoFoklis what can I say now

Well as you have mentioned the specific estimators predict better in most cases cos correlation coefficients = CCs are closest to 100% (or -100%).
Of course I can add more parameters in internals to improve the convergence and prediction but general approximation routines will not predict good for extrapolation cases.
I cannot present here more stuff about this cos I must post the source code (which is long) and it is hard to explain what is does or not.
Keep in mind that you have to use the General CPU speed estimator in the beggining and then you use the specific estimators.

Total AMD temperature analysis
The photo shows you the points which used in AMD gain and AMD temperature crash estimators.

Correlation coefficient: 89.7%

Model function: Exponential type

Samples: XP, Bartons, 754 and 939 single core (all desktop).

Cooling systems: Air, water and sometimes AC compressor.

Temperature range: +8.5 up to 59 Celsius.

Here the analysis stops.
More stuff after some years, I must collect data!
If you want to ask me feel free.

**v0dka** · 10-11-2006, 11:40 PM

Looks nice, what did you actually use to draw these graphs?

**MrSeanKon** · 10-12-2006, 12:29 AM

Thanks v0dka

All plots (black background) are from OcBible.
The others --> Mathematica and Maple sometimes.
To be more specific some AMD estimators use long formulas which are computed using Mathematica.
Moreover I must compare some OcBible results that's why these packages are installed also.
OcBible can plot more than one functions (up to four) together but this feature will available from v1.49.
Moreover I am thinking to add more capabilities just to gain time when we use these routines.
Someone may say "why use the OcBible instead of Mathematica and Maple???"
The reason is simple:
Both programs do not have a graphical interface you must type long commands such as Fit[data,Method->...] etc.....

**SoF** · 10-23-2006, 11:03 AM

I luv this math related stuph here

All estimators are based on watercooled and aircooled systems.
Someone (lots of people here) who have subzero what he/she should do??
I know that you don't have a subzero but I am wondering how can I/we get the most of this program.
I mean I know that predictions sometimes are rather correct as I used former versions of the bible for some own testing, but did you do anything for this in addition?
iirc some things were not able to work with negative temps(don't ask me which one it's way too long ago

), or am I remembering wrong?

**zipp0r** · 10-23-2006, 05:06 PM

i havent a clue what hes on about!!!

**TMM** · 10-23-2006, 10:56 PM

Originally Posted by zipp0r

i havent a clue what hes on about!!!

Haha indeed. Very mathematical approach there though MrSeanKon

.. i tend to be lazy and just take the "overclock until it bleeds" approach

**MrSeanKon** · 10-26-2006, 10:46 PM

Originally Posted by TMM

Very mathematical approach there though MrSeanKon

Thanks for comment.

Originally Posted by TMM

I tend to be lazy and just take the "overclock until it bleeds" approach

I knew that from your post in Bible's thread.

Originally Posted by SoF

I luv this math related stuph here

GKR I will BAN you GKR.....
You must use Maths GKR you are my friend

Originally Posted by SoF

All estimators are based on watercooled and aircooled systems.

Yes you are right.
Temperatures were from +8.5C up to 59 Celsius (if I remember well anyway this is shown in my personal AMD database).

Originally Posted by SoF

ISomeone (lots of people here) who have subzero what he/she should do??

Hopefully I found a sample at Subculture101 forum HIER CLICKEN/click here it was a FX-55.

Originally Posted by Alec at Subculture101

3GHz stable 24/7 @ 1.475vcore for superpi, prime, cs source etc (I am at 3GHz now).
3.2GHz bench stable @ 1.55vcore (might be stable lower aint really tried).

Thanks for the first sample!

Restrictions:
1) Alec did not say there if he uses multimeter for measuring Vcores at full load.
You must realize that all estimators are very sensitive.
Thus I assume --> Vcores were accurate....
2) Alec was not sure for the second MSF as you read there.
3) The subzero temperature at full load was -13 Celsius.
Mainboard showed how????????
I do not know......
Cos AMD gain estimator (in internals) uses the temperature difference.
Normally Alec must measure (with thermistor or an accurate sensor) two temperatures --> see AMD gain estimator help file.
4) I don't have any FX sample but Alec's CPU was 939 socket --> similar with my 939 samples (Winnie, Venice and San Diego).

Anyway let's play although I am not well informed!!!!!!!
The screen capture (Greek version) shows for upper MSF2 = 3170 MHz (68.8% possibility) which is near to 3.2GHz!
Alec said that the CPU is not stable at 3.2GHz (only for benchmarks).

Notes:
1) The second combobox has -13 for value but if you "pull down" it the range is 15.5 up to 52.5 degrees.
I entered -13 C --> cheating the OcBible --> Guidemania --> Top o/c guide --> Part 10 for more info.
2) -13 C differs enough from lower limit = +8.5 C --> high extrapolation --> danger of high errors --> hopefully program predicts well!

Guidelines to SoFoklis:
1) Set terrible and f*cking RAM timings such as 3-4-4-9 and 2T, and RAM divider just to ensure RAM stability.
2) Find the MSF1 setting Vcore = 1.45 - 1.5 Volts (it depends on your mainboard's BIOS settings).
3) Similarly find the MSF2 setting Vcore2 = Vcore1 + 0.05 up to 0.1 Volts.
4) For stability test choose the same program --> follow the Mapping method --> 8M/16M SuperPi, S&M quick FPU test or Prime/OCCT up to 20 minutes.
5) Post also the temperatures at full load.

Just to have a second subzero sample --> you will help me to post here more stuff about subzero o/c using the OcBible.

**SotosLg** · 11-20-2006, 02:07 AM

Sean Alec's CPU is an AMD FX not a single core like your samples. Since o/c exponent is around 0.7 for single core 939s why you calculate the MSFs using 0.6??? Is it wrong????

**MrSeanKon** · 11-22-2006, 01:22 AM