How to set up GTL Ref Values for 45nm & 65nm

[CryptiK]

I will try to explain. As Everybody has stated, the Intel specs say that quads have a reference GTL of 66.7% or 67%(hence the 0.667 or 0.67) of VTT/vFSB. But for some reason, even if the board gives the GTL voltages as per Intel's reference numbers, the chips are not stable.I don't want to go into "Who's fault is it anyway?" game between the mobo manufacturers or Intel. Remember, this only happens @ high FSBs with the quad.So to make the chips to be stable you think it's the Vcore and increase the Vcore, but still end up with an unstable chip and also unnecessary heat. That's the reason a lot of us with our experience with playing around with GTLs can give you a reference point like saying the 45nm quads seem to be stable with less of a GTL than the refernce of 67% of VTT.

Now, going into the calculation, let's say your rig is semi stable with a VTT of 1.30V. According to the reference from Intel, if you leave the GTLs @ auto, it will give the GTLs exactly that(0.667 times 1.3). But, for 45nms you want your GTLs lower than that. So you determine that you have to give negative values for the GTL so that board will give 0.667 times VTT-xx where -xx is the amount that you actually set in the BIOS of different boards. That's why when you see in my spreadsheet, I had put Auto as zero. Now the question is what does the -xx amount that you would have to specify in the BIOS. It's a simple math calculation. If you want to give a GTL of say 0.657, you are reducing the reference GTL by -0.01. So for a VTT of 1.30V, the equation becomes 1.3x-0.01x1000 = -13mV. The x1000 is because it's in milli volts.For a GTL for 0.647, you go 1.3x-0.02x1000 and so on and so forth.


Sorry about the long post. Hope this helps.

The thing is boards do not give just 0.667x for all GTL Ref's when left on auto. They give 0.635x for GTL Ref (0/2) and 0.667x for GTL Ref (1/3).

There are a few mathematical ways of working this out, just do whatever works for you. Your equation is fine. I gave the equation in the format in the first post as that is what is given in some bioses. When you just have the multipliers (0.635x, 0.667x etc), just choose whichever one you want and you're done.

This is a hard thing to explain, understanding how AGTL+ signals work can be tricky, but tuning them is the easy part, but for an inexperienced overclocker, it is something best left alone. To understand them properly you just need to do a bit of reading first.

In most cases for average overclocks, you are fine to just leave them on auto, and as always, if you don't understand what a setting in your BIOS means, don't touch it.

[Dee]

I got the same options in my bios but how the hell did you find out what your default GTL values are? And how high your VTT really is?

The BIOS tells me what the default GTL values are. I'm not sure about what the "real" vTT is, I can only go on what I set in the BIOS to be somewhere near accurate.

[Pisklink]

@ Dee

How does it report that to you, through hardware monitor or something? Now I'm curious if I could assume if my board has the same default values as yours?

[Dee]

My BIOS gives a brief explanation for each setting in a little pane on the right side of the screen. It tells you basic info, like the standard voltage a certain option is set to, or how raising a certain value may allow better overclocking... that kind of thing. Next to the GTL settings it tells me that one is set to 0.63x and the other is set to 0.67x when on Auto.

[Dee]

Here you go. Sorry about the blur, camera is a bit crap.

[Pisklink]

Thanks for the screenie!!!! I got the exact same Bios layout as I can see from the picture. Ofcourse I noticed this pane on the right side but never noticed this standard value. When I'm home I will take a look if both standard GTL settings differ for my board too.

[Hamilton]

would have been some great info if only DFI had defined the GTLs as a multiplier of vTT instead of the values they're using which has no relation to a multiplier or voltage...

[Baleful]

I'm still a bit confused on all of this, but from what I get from everything here's how I set mine:

FSB Voltage: 1.50v
Processor: Q6700
Motherboard: Evga 750i FTW

Equation:

1.50v x 0.667=1.50 x 0.635 + Y
1.001v=0.953v + Y
Y = 1.001v - 0.953v
Y = 0.048v
Y = 48mv

From this, I set all GTL Ref's to the closest number to 48mv, which was 50mv for my board. Presuming I did the math correct, is this right?

[davidk21770]

would have been some great info if only DFI had defined the GTLs as a multiplier of vTT instead of the values they're using which has no relation to a multiplier or voltage...

You have to go to the spreadsheet that they link to in the BIOS tuning manual and then convert in a much more complex way than has been discussed here! A real PITA and I'm an engineer (the stuff here is is really basic for me)! It wouldn't be bad if the spreadsheet had equations in it, but someone has replaced the equations with just numbers so you don't have the underlying formulas and you have to do intermediate extrapolations from their subtables. At least they could have left the formulas. In fact, they could have had it just compute the numbers for you! Who ever did the original sheet had everything needed.

[CryptiK]

would have been some great info if only DFI had defined the GTLs as a multiplier of vTT instead of the values they're using which has no relation to a multiplier or voltage...

As stated by davidk21770, refer to the linked table. I have no idea why companies seem to want to make it harder for you to work out. I have tried to keep it as simple as humanly possible in this thread.

I'm still a bit confused on all of this, but from what I get from everything here's how I set mine:

FSB Voltage: 1.50v
Processor: Q6700
Motherboard: Evga 750i FTW

Equation:

1.50v x 0.667=1.50 x 0.635 + Y
1.001v=0.953v + Y
Y = 1.001v - 0.953v
Y = 0.048v
Y = 48mv

From this, I set all GTL Ref's to the closest number to 48mv, which was 50mv for my board. Presuming I did the math correct, is this right?

Yes, you are correct

[davidk21770]

Generally higher vcore, and higher VTT you can run lower gtlref multipliers depending on clock frequency and multi for cpu. Setting the right GTL reference voltages is all just trial and error. Adjust VTT and adjust the GTLREF multiplier opposite of each other is what I usually do. If I modify VCore without modifying VTT then I adjust the GTLREF multiplier slightly higher, where as If I modify VTT without modifying VCore I will adjust it lower. Same goes for NB voltage, if the bios has NB GTLREF multiplier adjustment, you need to fine tune it when you modify vNB or vTT.

One other thing that I'll mention is. I always keep vTT higher than vCore. This might sound weird to most but it is actually how vTT is designed to work. It acts as a buffer for vCore during overshoot mainly during power on. If it is actually lower than vCore then there is a higher risk you can damage a CPU from too high vCore. It is also basically a limiter for valid BCLK high. IE. I run 1.43v vCore on my Q6600 G0 and 1.42V vTT in my Rampage bios (which is actually 1.47V vTT when loadline is enabled on my particular board).

I went in and decoded the DFI spreadsheet to guess at some GTL values for my MB (see sig) and my system is definitely stabler on that next OC that I'm working towards.

BUT, if I increase VTT to even close to Vcore, I don't even post. Are you suggesting that I need to reduce GTL to get the same thresholds that I have for the working VTT? Do you have an example?

[mikeyakame]

SOmething along those lines. If you change the vTT voltage and use the same GTLREF multiplier you would be increasing GTLREF also since it follows vTT.

If you are using 0.67x multiplier for example or 67% of a 1.30v vTT, that gives a GTLREF voltage of 0.87v. Lets say you increase vTT to 1.60v, since GTLREF follows vTT, you would now have a GTLREF voltage of ( 1.60 * 0.67 ) or 1.072v. The problem with this is that GTL reference voltage needs to be generally between 0.8-1.0v max. ANy higher and you generally end up with a situation where the NB can't figure out what is and what isn't a valid clock wave. In other words the end result worst case scenario is, no valid clock wave = no post since the NB doesn't understand what is and what isn't a clock signal.

So say I was aiming around 0.95v GTLREF at 1.60v vTT, I would need to reduce the Multiplier down to 0.59x or decrease the voltage by 0.12v or -120mV to get back to that point. This is just an imaginary value and generally you wouldn't need 1.60v vTT unless you were using over 1.60v vCore and 1.60v vNB.

THe idea of adjustable GTL Ref voltage is to be able to find a fine balance between the relationship of vTT, vCC / vCore, vNB, and most importantly to be able to raise or lower the voltage ranges where a valid clock wave occurs for both logical high and logical low waveforms. If you don't understand those it's not all that important, just remember that clock waveforms occur when you apply power to the circuit, then one clock later when you switch the circuit back to ground. Sometimes theres jitter, noise, cross chatter, and other electrically related phenomenons that disrupt or deform the clock waveform. Changing the voltage at which the reference for all other voltages in the system follow can move the clocks waveform into a region where previously the clocks waveform deviated from its wave path, you move the points at which the readings are taken. It's something like telling day and night apart during long days. The only way to differentiate is by the sun and moon but since there is heavy cloud where you are standing and you have no watch for all you know it might be night even though it looks like day. You move yourself from where you are to where the clouds arent, and doing this you know if its the sun or moon in the sky, thus if its night or day. GTL reference voltage adjustment isn't all that more complex than that for somebody who doesn't need to understand all the science and maths behind it. Moving it enough sometimes you get a clear enough sky to operate normally!

[Pisklink]

So say I was aiming around 0.95v GTLREF at 1.60v vTT, I would need to reduce the Multiplier down to 0.59x or decrease the voltage by 0.12v or -120mV to get back to that point. This is just an imaginary value and generally you wouldn't need 1.60v vTT unless you were using over 1.60v vCore and 1.60v vNB.

Are you saying that VTT should be as high as the lowest setting. In my case for example, I'm using the following in bios settings:

- 460 x 9.5 = 4.3Ghz
- CPU 1.4v
- CPU PLL 1.58v
- VTT 1.3v
- vNB 1.55v

With these settings the CPU was stable for 9 hours in othos. But when I increased RAM speed to 1840Mhz I thought it needs more vNB and increased it to 1.57v and 1.61v but now it fails after an hour. Could this have something to do with a low VTT?

[mikeyakame]

Well Intel spec says it should always be greater than vCC / vCore but in real world this isn't always necessary. From what I can figure out from the little actual explanation Intel gives about relation of vTT to vCC, it appears that in all cases vCC tracks vTT, where vTT is -+0.10v of vCC for that chip. Possibly vCC tracking vTT is for something along the lines of additional noise immunity for the signal, the tracking effect would limit the electrical potential between vCC and vTT, and if I understand this correctly the smaller the potential difference that exists the greater the effect of vTT being able to shield the vCC signal from noise which becomes somewhat critical as frequency increases. The immunity also works both ways, limiting the amount of noise leaked into other signals.

[mikeyakame]

Are you saying that VTT should be as high as the lowest setting. In my case for example, I'm using the following in bios settings:

- 460 x 9.5 = 4.3Ghz
- CPU 1.4v
- CPU PLL 1.58v
- VTT 1.3v
- vNB 1.55v

With these settings the CPU was stable for 9 hours in othos. But when I increased RAM speed to 1840Mhz I thought it needs more vNB and increased it to 1.57v and 1.61v but now it fails after an hour. Could this have something to do with a low VTT?

Adjusting the NB voltage close to CPUPLL can potentially result in leakage into the closed internal PLL voltage. The larger the gap between CPU PLL and any near by voltages the better usually. It's extremely touchy when it comes to noise or leakage and will destabilize the CPU eventually. If you raise NB to 1.61, take CPU PLL to around 1.68V at the sensor, vTT to 1.38 - 1.42 at the sensor, NB GTLRef if you have a multiplier aim around 0.56-0.62, and then you might be ableto bring vNB to around 1.57v. That is what I would start doing, and see what response I get and play from there. It's not scientific calculation or anything like that! Those voltages all respond in some what of a particular manner when taken that high. It'll be hard to stabilize the system your best bet might be to drop the performance level by 1, and bring the NB voltage down a little, the farther the gap between the NB voltage and vCC and the higher vNB is, the more difficult it becomes to keep the whole system from cross chatting and phases on data/clock/address signals keeping aligned with internal skew adjustment values.

The RF I have behaves similarly too, as have other Asus X48 boards I've played with. It becomes too difficult at a point where vNB is really high to compensate across the other values to keep the system from doing random and odd things. 1.59-1.61v NB on mine I found was about the limit before random behaviour plagued the system, cross chatter or noise related. This was only with 1.50v vTT, and I didnt really want to push it much higher since I didn't have enough headroom on vCC to keep going. That was PL6 pulled in to PL5 on phases a1/a3/b1/b3,333 strap and 1160mhz ddr2, and timings very tight, latency around 48ns.

[cadaveca]

Mike, i just wanted to post and say thanks for taking the time to explain this to others so that they may benefit.

Top Class, for sure!

[CryptiK]

Excellent info Mike, thanks, that brings a lot of what I have read elsewhere in bits and pieces together more cohesively. It sure is nice to have an trained engineer around

[Pisklink]

Mike thank you so much for your explaination. I read quiet a lot with regard to GTL on the techrepository.com and edgeofstability.com websites. However its described very theoretical and its in English which makes it hard to follow when it's not your native language. I think I understand you for the biggest part.

But I thought I read somewhere that the harm VTT can cause is very underestimated and that it should be treated with care because it can seriously damage your CPU. That is why I'm somewhat suprised to hear you say that it may be as high as vcore. I think there are only a few that use this knowledge when you would ask people here on the forum if there VTT value is nearly the same as there vCC.

At this particular moment I have the following OC result which is 4 hours stable on 4.4Ghz.

- Vid = 1.2375
- Strap 333Mhz
- Vcore = 1.4125v in bios
- CPU PLL = 1.58v
- VTT = 1.28v
- vNB = 1.55v
- LLC = auto
- Spread Spectrums disabled
- Dynamic read / write disabled



Now I want to increase RAM speed to 1860Mhz. I always thought that it would need extra vNB to reach that speed. But this is wrong according to your explanation because the gap would be larger between vNB and vCC. Moreover when I start finetuning vCC eventually the gap would become even bigger. I think I'm not able to run RAM at 1860Mhz with above settings (I will try ofcourse). Should I try lowering PL or should I add GTL and try to get voltages at the required distance from eachother to prevent destabilizing?

[dionysio]

better to use dual cores with maximus formula / rampage formula because of the gtl adjustment overlook... what is your opinion on this?

[mikeyakame]

i use a q6600 on my RF and its perfectly fine, i run it at 3.45ghz for every day stuff with 430MHz FSB and 8x multiplier. I never went to a 45nm quad core chip because personally the benefits of sse4.1 can't be utilized all that much yet since not much code has asm optimizations added to use the extra registers. I run 1.43-1.44v through it 24/7 and it hasn't missed a beat for the last 15 or 16months. If you want really high clocks definitely go for a dual core, but I've found a quad core running at 3.4-3.6ghz is much faster for compiling, encoding and even simultaneous drive defragmenting than a dual core even running at > 4.5GHz could ever be. In other words if you don't need the full use of a quad core, you will probably see more benefits from a higher clocked dual core. It has nothing really to do with gtl or anything else, the VRM on the MF/RF struggles at 500MHz FSB for a quad core, so if 500Mhz FSB and up is your goal then either of those boards aren't suited.

[davidk21770]

...but I've found a quad core running at 3.4-3.6ghz is much faster for compiling, encoding and even simultaneous drive defragmenting than a dual core even running at > 4.5GHz could ever be.

In other words if you don't need the full use of a quad core, you will probably see more benefits from a higher clocked dual core. It has nothing really to do with gtl or anything else, the VRM on the MF/RF struggles at 500MHz FSB for a quad core, so if 500Mhz FSB and up is your goal then either of those boards aren't suited.

The 1st quoted sentence seems to be totally the opposite of the next sentence?

(I have no experience with Intel duals -- went from an aging abused AMD 4800 to the quad in my sig. Just looking to see if I made the right choice).

[dionysio]

right now i am trying to get e8600 over 5ghz with rampage formula. duals over 5ghz beat quads at 4ghz in most apps except maybe encoding and like stuff. its 25% higher clock rate

asus x38/x48 are not capable to adjust gtl independently on each core pair.

[mikeyakame]

The 1st quoted sentence seems to be totally the opposite of the next sentence?

(I have no experience with Intel duals -- went from an aging abused AMD 4800 to the quad in my sig. Just looking to see if I made the right choice).

How so? I read over it again and it sounds right to me! For those who don't need to thrash the CPU for intensive tasks a quad core might not make that much sense. I personally find it does.

[mikeyakame]

right now i am trying to get e8600 over 5ghz with rampage formula. duals over 5ghz beat quads at 4ghz in most apps except maybe encoding and like stuff. its 25% higher clock rate

asus x38/x48 are not capable to adjust gtl independently on each core pair.

\Those clocks are too high to ensure data validity for anything extremely intensive. Clocks mean nothing if you cant ensure valid data which is critical for certain tasks.

And yep, no individual address / data gtl reference voltage adjustment for either die on the x38/x48 bios' which is kind of a let down.

[Pisklink]

The 1st quoted sentence seems to be totally the opposite of the next sentence?

(I have no experience with Intel duals -- went from an aging abused AMD 4800 to the quad in my sig. Just looking to see if I made the right choice).

Nah, I think it's not the opposite because Mike stats that when you don't use the quad to its full potential a dual will be faster. Meaning if you don't use it for encoding and stuff you would be fine with a dual core CPU. At least that's how I interpreted it.