• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

HHT/multi/divider formula etc. please help

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

Trypt

Member
Joined
Aug 27, 2003
Location
Mississauga, Ontario
I just have a couple quick questions.

Firstly, is it wise to go above 1000 hypertransport? The rig in questions is the s939 in the sig, and I'm wondering whether it's better to keep the HHT under 1000 at all costs or not.

For example, right now I don't need too much overclock, so I'm running at 250x4 for exactly 1000 (instead of 200x5 default), this is still a significant overclock over the stock Opteron 165, but the CPU can handle much more, I've had it over 2.5GHz stable no problem.

Now, if I want to run 260fsb, can I keep the HHT multi at x4 or is that going to cause problems? So first question is, is there any point in going over 1000, and how low can I go without it causing ANY slowing of the system.

I always thought that the hypertransport is way above what the system needs, so if I ran my 250x4 (as I do now) at 250x3 instead, for 750 instead of 1000, there would be no difference at all, since there is still much overhead, and in reality all the HHT needs to do is run at 400MHz to allow full throttle (unless memory is overclocked over 400MHz of course). Am I right on this, and if not, what is the lowest HHT that I should go to.

I'd like to run something like 280x8.5, but I fear that is way too high for x4 multi on the HHT, there is no 3.5, so I'm stuck with 3x which well under 1000, 840 in fact. So, what is the lowest?

This brings me to the secondly. Is there an exact formula for CPU to memory multiplier. It's incredible how crazy the memory MHz numbers get when I change the multiplyer. If I keep it at it's highest 9x, then it's just a matter of CPU/11, but as soon as I change the multiplyer, everything goes haywire, for example if I run 300x8 for 2.4GHz, the usual formula which works with 9x multi doesn't work and the memory is trying to run at over 300MHz (600Mhz DDR), craziness. One time I tried 320x7.5 and when I put the memory multi to it's lowest (1/2, 100MHz), it ran at over double it's rated capacity (almost 400MHz/800MHz and failed immediately), and funnily enough, when I put the divider on the memory higher, the frequency went lower, and changing the cpu multiplyer changes it all again. It is only clear to me when I keep the multiplyer at 9x, but that doesn't allow me to achieve my full O/C potential, unless of course my first inquiry here is correct and HHT really doesn't matter and can be run as low as 750MHz without any loss of performance (750MHz is a number I'll never run that low, it's just an example from above, but I would like to run 280x3 for example for HHT, and a 8.5 multiplyer, but again, that screwes with the memory and I have to test every setting and even then I have no guarantees, I need a formula if there is one. I am an advanced computer user, builder, repairer, overclocker etc, but this is eluding me, but I am an astrophysicist, so any formula will be a big help no matter the complexity.

So, is there a formula that takes into account FSB and the multiplyer of the CPU and the memory divider together on the Ultra-D s939 motherboard?

Thank you.
 
OP
Trypt

Trypt

Member
Joined
Aug 27, 2003
Location
Mississauga, Ontario
While I would love the formula, if indeed the HHT doesn't matter and if you tell me that lowering the multi on it to 3x from the 4x I have it at now, from 1000 to 750 would make absolutely no difference, that would fix my problem anyway because then I could keep my CPU multi at 9x and just increase the FSB from 250 up and down the HHT multi to 3x even though it would be significantly lower then 1000 (but higher then 800).
 

RGone

Senior DFI Staff
Joined
Apr 20, 2004
Location
JAX, Mississauna
During testing back in the old days of those chipsets, it was found that underclocking HT a couple of hundred Mhz could result in a performance gain/loss of +0.1% to a -0.6%. However an overclock of the HT by a couple of hundred Mhz would result in no gain but all losses from -0.1% to -4.5%. You can make up your mind about your situation from that data.
 
OP
Trypt

Trypt

Member
Joined
Aug 27, 2003
Location
Mississauga, Ontario
Thank you.

Now, about the formula.

Anyone know how it changes depending on the cpu multiplyer used?

If I use the stock 9x multi on my Opti 165, the formula is simple, the higher the fsb, the lower the divider on the memory in order to keep it within an overclocking range the mem can handle (in my case, under 220MHz or so).

But if I change the multiplyer to 8.5 or 8 or anything else, in order to get my FSB high and use the x3 HHT multiplyer (to get as close as possible to 1000, for example 315x3 for 945), I can't use 9x multi because that is way too high for my CPU to handle (315x9 is too much for my Opti, WAY too much.., so I have to lower the cpu multi), and when I do that, the divider memory stuff goes to crap, nothing is as it should be, it goes all over the place, I can't figure it out, it's like it has a mind of it's own, sometimes a higher divider giver a lower clock then a lower divider, it's just weird.

I wonder if there is a formula(s) from "those days" that explains how it works when messing with both FSB and the cpu divider to figure out exactly what memory divider to use for a desired frequency.

As it is, right now at this momeny I'm running 263x9 for 2368MHz, just to see how it responds to a low HHT (263x3HHT) and it's much faster then even my 250x4HHT and 2250MHz, so 118MHz overclock over the old one, but an HHT much lower then 1000 even is much faster, weird.

Thanx anyway.
 
OP
Trypt

Trypt

Member
Joined
Aug 27, 2003
Location
Mississauga, Ontario
Awesome, thank you so much, that is EXACTLY what I needed. An online calc is even better then a formula, lol, although I can deduce that from the calculator itself, which I have already done, thank you again.
 
Last edited:

RGone

Senior DFI Staff
Joined
Apr 20, 2004
Location
JAX, Mississauna
So many of the 'good' links that had spread sheets etc are all getting very old and are missing from the 'text' in the various forums. But I located that and did the math with it and it kept with the obtuse formulas so I just linked it. Good luck man. RGone...ster.
 
OP
Trypt

Trypt

Member
Joined
Aug 27, 2003
Location
Mississauga, Ontario
It's actually not totally accurate, it is accurate with certain multiplyers, but not with others, oh well, I'm only interested in 8x, and it's off by 5% which is acceptable, and I have already incorporated that error into my calculations. I run my system at 303x8 now for 2.425GHz and the ram at 203MHz. The formula says the ram should be at 209MHz, but I have ran so many tests with ram and the performance between 205 and 215 for example is completely negligeable, much less then 1%, so I just don't care, as long as the memory is not under 200MHz I'm ok with it. I tried the higher divider and the ram runs at 218MHz then, but with that speed, it fails Prime95 after 9 hours on the middle setting (large FTT's, some memory tested) on one core (the other core is extremely stable for some reason, it never fails no matter what, even at 3GHz, but the second core fails very quickly at some settings, so this current setting is the highest at 1.35V (cpu voltage) that the second core will go (the first core will run 2.4GHz at lower voltage then it requires, lol, it must have been one hell of a core, mixed with a crappy core).

Funnily enough, when I run both cores on P95, each instance running a core, the second core fails faster, when I run only the "bad" core, it doesn't fail at all sometimes, sometimes after 10 hours etc, so it must have something to do with cooling, not just voltage, because when I run only the "bad" core on P95, the temp is much lower then if I run both of them each with their own P95.

So I think the system would be stable no matter what, but I want it completely stable in P95 and SuperPI (btw, it passes SuperPI no problem at the above setting, 32M in 33 minutes, which is about 10 minutes faster then stock speeds, and 4-5 minutes faster then my usual run of the mill overclock of 250x4HTT and x9 multi for 2.25GHz.)