The goal: FSB>475MHz
Last night, I was playing with some settings in the BIOS, and was successfully able to get my PC to POST and run memtest86+ at FSB speeds of up to 510MHz. The problem is, as soon as I start trying to actually boot windows, I get lockups above 475. Just need to figure out what's getting unstable!
0
Comments
I'm actually not quite sure what GTL reference voltage is.
Edit - I think it's actually named differently in the Gigabyte BIOS. What you refer to as GTL reference is probably CPU reference or northbridge reference or something like that. There are options for reference voltages for 3 or 4 different things in the MIT settings.
A good read and some good links in the writeup. IMO, your at the stage where you'll have to start tweaking the reference voltages. If it's the rig listed in your profile, I don't think the 8 gigs of ram are doing you any favors.
You could also drop the multi of the CPU, and see if you hit a roadblock upping just the fsb.
Intel chipsets use a type of voltage signaling called A/GTL+, or Assisted Gunning Transceiver Logic+. A/GTL+ is a set of reference voltage levels that the CPU uses to determine whether a signal is high (binary 1) or low (binary 0).
Precise circuitry on every Intel motherboard regulates this reference voltage as a percentage of the vTT, or what systems with BIOS vTT control often call the "FSB termination voltage" or "vTerm."
The standard A/GTL+ voltage swing is between 0.4v and 1.2v, with the reference voltage pegged at 0.8v. If the signal rises above 0.8v and falls within or above the rising noise margin (รขโฐยฅ0.9v), it's a binary 1. If it ducks below 0.8v and falls within or below the falling noise margin (roughly รขโฐยค0.7v), it's a 0.
NOTE: The noise margin is an arbitrary voltage threshold that's supposed to filter out ringback.
Now we need to discuss that ringback phenomenon. When the FSB's signal is rising from low voltage to high voltage, or falling from high to low, we can say that it is rising to the pull-up voltage or falling to the pull-down voltage, respectively.
In the case of rising to the pull-up: The FSB signal has a tendency to overshoot its target voltage and then strongly dip back under the target before resting in the settling limit. This strong dip under the target is ringback.
In the case of falling to the pull-down: The FSB signal can overshoot the target voltage (counter-intuitively called undershoot, even though it's going over), and then bounce back up above the target before eventually settling out. This jump over the target is also called ringback.
Now, if a signal has ringback that crosses into the noise margins, or <i>especially</i> beyond the noise floor/ceiling, the A/GTL+ circuit could mistakenly interpret that as a 1 or 0: This will corrupt data on the bus by introducing data that didn't come from the input. Hello bluescreen! Hello boot failures! Hello hard locks!
Let's establish a few things:
1. The CPU and the northbridge operate on split power planes, which means that each voltage has its own GTLREFv multiplier.
2. The CPU's GTLREFv multiplier is usually 0.67x, or 2/3 the vTT.
3. The typical GTLREFv is 0.8v.
4. If the CPU reference threshold is 0.8v and that is 2/3 of the vTT, the default vTT must be 1.2v (it is).
5. The northbridge GTLREFv multiplier is usually 0.61x or 0.63x, or 61% and 63% of the vTT.
6. Using the same maths, we can then decipher that the MCH uses areference voltage of 0.75v or 0.73v.
7. Intel specifies a maximum vTT of 1.55v.
BASIC ADVICE:
Rising to the pullup target of 1.2v is harder for the signal to do than cutting power to the pulldown target. Ringback on a rising signal is usually the culprit of the blue screens, crashing and freezing people see when overclocking a quad into 500MHz territory.
Lowering the CPU & MCH Reference Voltages ever so slightly can keep the strong rising ringbacks on a high FSB from plunging beneath the noise floor to create false data on the bus.
The rising FSB clock has ringback below the noise floor at 0.8v GTFLREFv
Lowering the noise floor to 0.73v (GTFREFv 0.61x) keeps the strong ringback within tolerances
Always tweak the GTLREFv before touching the VTT. Because the VTT sets the high voltage point for the core components in the system, raising the VTT will make everything work harder to hit a higher voltage which means more heat and reduce component longevity. Also, raising the VTT without adjusting the GTL Reference multiplier will raise the reference voltage instead of lowering it as is often necessary.
I hope this has given you the appropriate insight on the issue.
Edit: I have CPU Reference, MCH Reference, and MCH/DRAM reference. Problem is, all of them seem to be expressed as actual voltages, not ratios. I tried messing with the CPU term voltage, but any setting other than "Normal" seems to cause a no-post at 500MHz FSB.
Second edit. Meant CPU REFERENCE voltage, not term voltage. I need a coffee, just got up.
The default MCH reference is around 0.73v.
You'll want to bring those down to around, say 0.73v (CPU) and 0.71v (MCH), if not slightly lower.
Also make sure that your FSB is using the 333 or 400MHz strap, and that you've boosted the vMCH to somewhere +0.15 or +0.2v range. I forget the default vMCH so I cannot give precise values for that.
Also loosen up your timings on the DRAM as far as they will go and use 2.1v for that.
Maybe I just don't have the patience for it. The intel matrix storage manager has been starting to detect errors on the raid1 boot array as well, meaning I'm gradually corrupting things on the drive with this experimentation.
I have determined my system to be absolutely stable @ 466MHz FSB (3.72GHz CPU) with 8GB of ram installed and very modest voltages. I've decided this is where I'll keep it, it should be my best bet for performance without killing the PC too quickly.
I dunno. In any case, as I mentioned I have to give my CPU too much voltage for my current cooling setup above 475, so I'm sticking with my 466 24/7 OC with 8GB ram.
http://www.bit-tech.net/hardware/memory/2008/07/08/is-more-memory-better/6
If its not a server, you would probley benefit from the extra fsb/clock speed over the extra 4gigs of ram that more than likely sit unused most of the time.
Yorkfield quads have a TjMax of 100C, I believe, so that would be load temp of 88C. 25C from TjMax is hot (75C) but within tolerable limits, as Tcase has not passed intel's spec max (70C) although it is probably close.
I know above 3.8 my q9550's were starting to take quite a bit more voltage for stability (and a bunch of tweaking) The only one I have left is sitting at 8.5x472 on a UD3R. I was pulling for you to hit 475fsb.
Do you have a fan blowing onto the NB and ram? It might help you some.
When my load temps hit the 65c range, I quit pushing and went water.
FWIW, I'm at 1.35volts +/- for 4.0, I believe it was 1.27 for 3.8