Athlon64 Overclocking Optimizer

[Gogar]

Interesting... so I noticed you use a Sum Square Error for your "objective function". Are you using an optimization algorithm or just searching the solution space for the maximum?

Edit: It seems like you're just displaying all solutions... are you weeding out any of them?

Just searching the space. It only consider the dividers and multipliers around their required values for a range of the HTT frequency.

I don't think you can even use an optimization algorithm because the surface of the memory speed is very irregular. If you gradually increase the cpu multiplier the memory speed will go up and down all the time, it's not a smooth surface at all, so most optimization algorithms would likely get stuck in local minima's.

[Ackbar]

Just searching the space. It only consider the dividers and multipliers around their required values for a range of the HTT frequency.

I don't think you can even use an optimization algorithm because the surface of the memory speed is very irregular. If you gradually increase the cpu multiplier the memory speed will go up and down all the time, it's not a smooth surface at all, so most optimization algorithms would likely get stuck in local minima's.

Interesting! I guess there aren't too many choices and they're all discrete so you don't have to worry about non-integer values. Very nice implementation!

[S!1v3rB@cK_Dk]

I still dont understand this error thingy! Is it good to have a less higher amount of errors? And what do they actually stand for?

[agenda2005]

Hey Gogar,
This is an excellent tool. Thanks for your time. I see that higher multi usually gives large error even when you run memory at 1:1. My system is SD 3700+ and prime 95 stable at 2700MHz at 246*11, memory top at 285MHz with 9 multi, but I would not use it past 277MHz for 24/7 operation, however my 10X multi is shaky, but the optimizer comes with best scenario for 10 multi which I also prefer.
Is there anyway you can factor in the multi as one of the important weights?

[GazC]

I was hoping it'd give me a combo that would let me get the last few stable Mhz out of my ram, but it seems the current settings I use for benching is the best one. It's a useful too though and will be very useful for everyone, thanks.

[Gogar]

I still dont understand this error thingy! Is it good to have a less higher amount of errors? And what do they actually stand for?

The lower the error, the better.
It's a weighted sum of the squared percentage it's below the maximum. And yeah, it's ok to think "wtf?"

The error value doesn't mean much at all by itself, just their relative sizes are necessary to order the table and put the best one at the top.
To see how good an entry is it's still easier to look at the other columns to see how many MHz it's below the maximum. I can just as well leave out the entire error column. Hope this helps

[Gogar]

Hey Gogar,
Is there anyway you can factor in the multi as one of the important weights?

Actually, yes.. Indirectly through the HTT.

If you tell it to favour a high HTT, a low multiplier will be needed for that to make the cpu frequency.

Or do you want it to use higher multipliers?

[agenda2005]

Actually, yes.. Indirectly through the HTT.

If you tell it to favour a high HTT, a low multiplier will be needed for that to make the cpu frequency.

Thanks. Using HTT of 1080 and 90% priority, I figured that 271X10 will be my best, but I need a BIOS that will stabilize this 10X multi. 271 X 10 gave an error of 7.2 compared to the others that are 45 and up to 321. My stable HTT is 1080, RAM is 285 and CPU 2710. Good job m8!

I was hoping it'd give me a combo that would let me get the last few stable Mhz out of my ram, but it seems the current settings I use for benching is the best one. It's a useful too though and will be very useful for everyone, thanks.

Not so fast m8 . The purpose of the optimizer is for you to figure out your max for all the parameters

RAM/HTT
CPU
Multi
HT
and then priortize them according to what you feel is more important to you.
For example if you want to run an application that is bandwith intensive. You need a higher HTT/RAM Speed. But you might also want the HT (Hypertansport) speed to be as high as you can. The question this optimizer will answer is if you are better of running a divider on your ram or not and what will using half multi will have on performance. You need to isolate each componet, stress them out and priotize them in percentage and let the optimize do the rest. This is a good tool m8.

[GazC]

Not so fast m8 . The purpose of the optimizer is for you to figure out your max for all the parameters

RAM/HTT
CPU
Multi
HT
and then priortize them according to what you feel is more important to you.
For example if you want to run an application that is bandwith intensive. You need a higher HTT/RAM Speed. But you might also want the HT (Hypertansport) speed to be as high as you can. The question this optimizer will answer is if you are better of running a divider on your ram or not and what will using half multi will have on performance. You need to isolate each componet, stress them out and priotize them in percentage and let the optimize do the rest. This is a good tool m8.

Erm, yeah I know. All I was saying is that the best combination it suggested was the one I was running, what is wrong with me saying that?

[Pjotr]

HAH! All your Base if you are screwing up (I typed in 2.55 for the mhz instead of 2550)
I like it makes it so I don't have to think (which hurts at times)

If you had to enter 2.55 it would have said "GHz", now it says "MHz", I don't see how you can have to think *less*? Just enter what it says and don't think!

[Gogar]

Anybody got any ideas for new features? I'm kind of wondering if i can make it optimize for specific applications. Or even find you a set of hardware that will produce the highest scores for a specific application.

[I.L.F. 3004]

Nice job Gocar.. handige tool

[^don.k's^]

I wondered (some time ago) about some maths helps to maximize out my OC's, i couldn't do it, but you could! Very good work man!

[IYP]

[daddy_fizz]

so is there any way you could compile this like some other people were saying? would be nice to fiddle around with this while i'm offline...

thanks

~Fizz

[Gogar]

I have a calculus exam coming up soon (which i failed before coz.. err.... DOOM3 came out ) so i'm quite busy studying for about another two weeks, then i'll look into making it an app. (after June 14th)
I didn't think there would be any people who aren't online anymore these days though :-P

Edit; Just added a htt field to the memory divider table, check it out.

[MrSeanKon]

Look attach photo:

Of course I had a 3200+ 754 socket....
But I choose 3500+ cos of I set 11 multiplier.

I see best choice 9.5X260 --> 2470MHz (9th line)...
5th line shows 10.5X240 --> 2520MHz...

If I run my system as you suggest it will run slower.....
Cos my main applications (video encoding) needs FPU performance thus MHz speed is the main factor not FSB....
OK I know other applications (Maths + large arrays) which use too much memory and the total amount is changing during execution...

The important thing for optimizing any system (Intel, AMD , 286, 386 etc) is different in my opinion.
RAM Optimizer is available for more info...
Anyway if I think wrong please be specific.

[Gogar]

Of course I have a 3200+ 754 socket....
But I choose 3500+ cos of I set 11 multiplier.

That's good. Sorry that it assumes you have socket 939, but i wanted to make it easier to understand for people with socket 939 (most are, right?), and so i put 939 model numbers next to the multipliers.

If I run my system as you suggest it will run slower...

Well it's not surprising entries lower in the table are bad choices. You should always look at the top row of the table first, if you don't like that combination (or if it's impossible to use on your mainboard) then look at the second row etc.
If you look at the top row it says 10*255 = 2550 1:1(200) x 255 = 255
That's not a bad choice at all, your memory runs only 5MHz lower than it could while the cpu runs at max.
Perhaps you could try to get your cpu stable at 2600MHz with a small vcore bump, then your memory can run at 260

Cos my main applications (video encoding) needs FPU performance thus MHz speed is the main factor not FSB....

As long as FSB is set to "0" in "Importance Weighting" it's not part of the equation, in other words: it won't try to make the FSB high. (on A64 the FSB has NO influence on performance)

The important thing for optimizing any system (Intel, AMD , 286, 386 etc) is different in my opinion.

I think i know what you mean, but you can control it's behaviour with Importance Weighting.
If this were for Athlon-XP you would probably want to set FSB higher than 0 in Importance Weighting.

[MrSeanKon]

OK I understand cos I do not use RAM dividers.
If my RAM cannot run 1:1 I prefer to relax timings or FSB.
Your approach is different.
Thanks.

[agenda2005]

OK I understand cos I do not use RAM dividers.
If my RAM cannot run 1:1 I prefer to relax timings or FSB.
Your approach is different.
Thanks.

You still need to look cearfully at your table. It said your best combination is 255 X10 = 2500MHz with RAM at 1:1. This gives the lowest error, memory running at 1:1, you have to decide what timming is stable for your ram and you also get the highest possible CPU MHz that you asked for.

[ImportantAwareness89]

So you wrote this program having just taken Calculus. Brilliant. This uses some stats? Some least squares stuff? (Linear alg)... Wow.

*frowns at D+ in Diff eq./Linear alg.* so much for acing calculus.

[Gogar]

Added another page which lets you browse the frequencies, multipliers and dividers.

Just click on the select boxes, and it'll show all possibilities from there.
http://www.gogar.com/a64ocbrowser.html

[Gogar]

Added Rev E memory dividers, it was kind of a big change as some stuff was hardcoded so be careful with the results Everything seems fine so far, but i have no way of testing. Please let me know if anything is wrong with it.

Oh and another thing, i'm looking for bugs in the A64 itself.
If you look at the table
http://math.gogar.com/athlon64.cgi?showtable=1&htt=200

The fields in red are where the actual memory speed is different than what you'd expect.

I suspect some bugs might be among the new memory dividers as well, especially the 13:12 divider looks like a good candidate. But i have no way to test.. so if anyone finds any differences between the real memory frequency they get and what my table says, then please tell me about it.
Thanks

[red_dragon]

hmmm wow. thx I'll give it a try

[MrSeanKon]

It said your best combination is 255 X10 = 2500MHz with RAM at 1:1.

I knew this.

You still need to look cearfully at your table

OK I was not very carefull while I was looking this large table but I became confused cos of many choices.
I think no reason for second line (10 X 255 = 2550 MHz RAM 9:10) cos fourth line (11 X 231 = 2541 MHz RAM 1:1) is prefered by me.
As I said above I follow a different approach --> A64 o/c suggester thread is here at XS