(originally posted by R0b0tc0rpse)
 
This post will try and explain two things, FIRST, what is GTLREF, SECOND, how does it apply to our BIOS (7 Series Nforce). And do to it in a general way that the layperson can understand, so that you don't need to have a CE degree to read this.

Please read the whole thread before asking questions, your question may already have been answered. Ask questions in this thread, and I will try to answer them when I visit this thread, I will not answer questions sent via PM, or questions on general overclocking or how to use benchmark tools.

**Update for 790i owners
**Update for Dual Cores

I spent A LONG TIME compiling and a lot of hard tedious work testing this information, DO NOT copy paste any of this into your own posts in other threads or other forums, rather, post a link to this thread.

This is not a guide on how to overclock, this assumes you know a thing or two about overclocking and your rig is stable enough to boot to Windows and run prime95.

This is not going to give you a magic number that will make you stable. Significant effort on your part is required. Everyone's GTLREF Adjustments will be different; it's best to find your own rather than copy some one else's.  As little as 5-15 mV can mean the difference between prime95 stable for a few minutes and for days, well with in the variance you'd expect to find between motherboards/CPU'S of the same model.

What is GTLREF?

GTLREF is NOT Vcore adjustment for individual cores

GTL is defined by JEDEC standard JESD 8-3 (1993) and was invented by William Gunning while working for Xerox at the Palo Alto Research Center.* (Wikipedia)

AGTL+, or Advanced Gunning Transceiver Logic Plus is a derivite of GTL found on Intel based systems. Strictly speaking, Gunning Transceiver Logic is a way to filter signals being sent between microprocessors to eliminate electrical noise. The term GTLVREF refers to the reference Voltage of a GTL lane, hence GTLVREF (GTL Reference voltage). Essentially it sets where the noise margin is for a particular signal. It tells the microprocessor where to filter out the electrical noise in the signals associated with crosstalk and ringback, etc (primarily ringback). GTLREF voltages are supplied to the microprocessor on completely different lanes than the signal they are a reference for filtering, the voltage it self is not a signal and is constant.

There are plenty of technical articles on how it works if you want to get more in depth, but it requires some basic understanding of physics, calculus, electronics... etc. Just Google AGTL+ or Gunning Transceiver Logic.

Before motherboards started putting GTLREF adjustments in the BIOS, SERIOUS overclockers (not number pluggers) would solder potentiometers onto their motherboards to allow them to adjust this with the aid of voltmeters and oscilloscopes. There is a GTL for your CPU, memory, there is a GTL for the FSB/NB, there's GTL's for lots of parts of the motherboard. Most of these are not adjustable out side of making board modifications. A lot of motherboards (especially enthusiast ones) supporting Core 2 processors have SOME form of CPU GTLREF adjustment.

AGTL+ referance voltages on Core 2 processors are for filtering singals between cores on the same die.

I can spend 3-4 pages explaining how it works, getting very technical, but most people aren't interested in this... They are interested in how it effects them.

ALL MOTHERBOARDS ARE DIFFERENT; they don't always let you adjust CPU GTLREF in the same way. There are a lot of articles out there talking about GTLREF, and they each talk about a different motherboards. Most are "Self-Normalizing". Some motherboards allow you to set the RATIO of GTLREF:VTT, some let you set the total as a number (some in V, cV, mV). Some let you set the RATIO and the OFFSET. Some let set lanes individually, all at once, or in pairs. In the case of the 7 Series Nforce, it is "SELF-Normalizing", which means the motherboard calculates the CPU GTLREF voltage as Commanded VTT*.67, and allows you to adjust the 4 CPU (or 2 for dual cores) GTL lanes with additional voltage (an offset) from 0-160 mV in increments of 5mV. A setting of 0mV will result in the motherboard commanding COMMANDED VTT*.67 for each CPU lane.  *In the case of the 790i you can adjust from -155 mV to +155 mV in 5 mV increments.

Adjusting GTLREF:

DO NOT POST CHARTS OR INSTRUCTIONS FOR OTHER MOTHERBOARDS IN THIS THREAD.
It will only confuse other people.

**Note: From here on out, we are only talking about CPU GTLREF because that is the only GTL lanes you can adjust in the BIOS. Other lanes exist but require board modification to adjust.

**Note: A lot of motherboards, particularly those with Intel chip sets, show GTLREF totals in unmarked units ("GTL Units"), which are similar centi-Volts. These units, while similar to cV, scale with VTT, while 80 GTL units @ 1.2 VTT is equal to .80V, @ 1.3 VTT 80 GTL units are equal to .855V. DO NOT get confused by those they have NOTHING to do with the 7 Series. For example, in one of these motherboards, at VTT 1.2  a "GTLREF setting" of 80 would be equal to 0mV in a 7 Series. A setting of 90 in one to those motherboards would be a setting of 100mV in a 7 Series.  You will never see the total GTLREF voltage in a 7 Series motherboard in volts or otherwise. You will never use "GTL units" in this motherboard. 

In the case of the 7 Series Nforce motherboards you are given the ability to adjust, in increments of 5mV, from 0 - 160 mV four separate CPU GTLREF lanes. Please note, that this type of adjustment precludes the ability to ever make adjustments below what the motherboard is attempting to set. This value in the bios refers to additional voltage you can add to that lane, and not the total GTLREF voltage. The computer would not work at all with a total GTLREF voltage of 0. *In the case of the 790i you can adjust  
 
from -155 mV to +155 mV in 5 mV increments.

This means at a setting of 0mV GTLREF Offset adjustment in the BIOS will result in the motherboard using the "SELF-Normalized" value of [Commanded VTT*.67] for voltage on all four CPU GTLREF lanes. This may or may not be stable for your overclock.
 
VTT is "FSB Voltage" in 7 Series Motherboards.


These lanes available for adjustment in the 7 Series Nforce represent the following:

FOR QUAD CORES:
Lane 0 is for Core 0,1 Data Bus (die 0)
Lane 1 is for Core 2,3 Data Bus (die 1)
Lane 2 is for Core 0,1 Address Bus (die 0)
Lane 3 is for Core 2,3 Address Bus (die 1)

FOR DUAL CORES:
Lane 0 is for Core 0,1 Data Bus (die 0)
Lane 1 is for Core 0,1 Address Bus (die 0)

**Note for dual cores:
It's possible that, despite what I read in the published Intel articles that with a dual core, lane 2 in the BIOS could still be for the address bus of core 0-1. We would have to have Nvidia confirm that. I do not have a dual core to test with. Since we know with 100% certainty that lane 0 is still the core 0-1 data bus, go ahead and replicate your lane 1 setting to lane 2 and lane 3 just to be sure. From what I understood the GTLREF voltage circuits on the motherboard were all the same and each one simply sent voltage to a particular pin on the processor socket, and that on a dual core, the pins for lane 2 and 3 were not used and the pin for lane 1 became core 0-1 address bus. Any extra voltage on unused lanes would be harmless.

http://forums.evga.com/tm.asp?m=476249&mpage=1&key=񴷲
I have now confirmed (look for my reply in this thread) that it is the case that on a Dual Core (single die) processor the pins for lane 2 and 3 are not used, and that lane 1 becomes Core 0-1 address bus. On dual cores, 0 on lane 2 and 3 should suffice, but there may be stability issues with the board it self when there is great disparity between the settings of the four lanes, so setting them to what ever your lane 1 is would be fine. This is of coarse, if Nvidia didn't pull a programing trick and reassign the name in the BIOS when a dual core is detected. Either way setting lanes 1, 2, 3 the same on a dual core should work fine.


Some other basics you need to know to understand the rest of this post;
When I say COMMANDED voltage, this is the voltage value you are putting into the bios or the voltage the motherboard is trying to apply to a lane.
When I say ACTUAL voltage, this would be the voltage in reality if you put a voltmeter on that part of the motherboard, while this is almost always less than COMMANDED, there are instances where it can be more.
When I say NOMINAL GTLREF RATIO this would be the ratio that is actually stable for the current FSB speed/VTT voltage/Vcore, and not .67

Behind the scenes the motherboard will always use ("SELF-Normalize" to)[COMMANDED VTT*.67] for lane voltage, this means, if the VTT is 1.2 the motherboard will COMMAND GTLREF for ALL FOUR LANES of .804v, or 804mV.  This is what INTEL DESIGNED it to be. How ever they admit that this ratio will shift when changes are made to FSB speed, VTT voltage, and Vcore, and it can shift in either direction based on those changes. No company "supports" overclocking, but Intel, and other companies share their research, and we have found that increasing GTLREF can be necessary for stability while using increased VTT/FSB speed/Vcore. 
 
If you were to actually measure the GTLREF lanes when commanded VTT equals 1.2V you would likely find it somewhere between .750V and .810V
If you were to actually measure VTT at 1.2 commanded voltage, you would likely find it between 1.18V to 1.22V.
 
The 7 Series motherboard has no way to monitor actual VTT volatge or actual GTLREF lane voltage
There are two reasons to add voltage to the GTLREF voltages.
1: You measured a voltage drop and are compensating for it or:
2: You have increased FSB clock or VTT and/or vcore and the motherboard's normal GTLREF voltage
of (commanded VTT times .67) is not stable and you have to shift your noise margin to a
higher ratio   
 
The goal is to get all four lanes ACTUAL voltage to be the same, always. This is a bit difficult, but that is why they gave us the ability to adjust each lane.

Say we soldered on some wires to our motherboard and can measure ACTUAL VTT, and ACTUAL GTLREF lanes.

We would then, for each lane, take [(ACTUAL VTT V*.67)-ACTUAL GTLREF V]*1000 to find out how much adjustment we need to make in the BIOS.
IF for some reason you get a negative number, then you are sunk, because at that point the only way to do that would be to solder in a potentiometer and adjust it with a volt meter manually. Unless you have a 790i which allows negative adjustment.

If the value is between 0 and 160mV, then we are lucky because the board will let you do that. 
 
So you take this adjustment and round up to the nearest 5mV and would then apply that to the lane you had measured, repeating the whole process for each lane. You would then measure again to make sure it is correct and you'd be all set.

The address bus lanes are less prone to voltage drop because they are not as heavily loaded as the data bus lanes, so they usually require less of an adjustment.

If you were increasing VTT and or FSB speed and or Vcore, then you would likely hook up an oscilloscope while adjusting the FSB speed with a clockgen and GTLREF with a potentiometer to find your new noise margin.

But I don't think any one really want to solder wires onto their motherboard to find out the actual voltages, do you? 
 
Here is a method any one can do, with out needing to modify your motherboard:

Interestingly enough, I found that 1.2 FSB voltage can take you pretty far with correct GTLREF adjustment, and you should avoid increasing FSB voltage because that introduces more unwanted heat and signal noise. I have found my rig just as stable @3.6Ghz w/ 1.2 VTT with correct GTLREF as it was with 1.4 VTT. Right now I'm working on 3.84Ghz.

Try with 1.2 before moving incrementally up .1 at a time. But you'd be surprised what you can do with 1.2.

1.2 VTT should be good up to just over 400 MHZ FSB speed (1600 QDR) and for alot of us that should be plenty. Over 400 MHZ 1.3 will take you everywhere you want to go unless you are getting WELL over 4ghz with an 8x multiplier. Any one intrested in this range likely has a QX and would scale the multiplier rather than push the FSB much higher than 2000. I think it is better to run a higher multiplier and more Vcore than it is to run higher FSB/VTT which more greatly effect GTLREF adjustment. 1.4 and 1.5 will almost never be used unless you are well over 2000 QDR with a very low multiplier, or have significant voltage drop @ 1.3. High VTT's that people are use to seeing (1.4) has been an attempt to compensate for improper GTLREF adjustment, either from (auto) or manaully set.

Starting with all lanes set to 50mV (which is a good starting point for most overclocks) or better yet 0mV if you can boot that way. *790i owners with 45nm Dual cores may want to start off at ~-50 to ~-70 and work up from there. Especialy if you can run 1.1 VTT.

Boot into Windows (assuming you can boot into Windows) and start a prime95 torture test.
Watch the cores fail...
For quad cores:
When the failure is on core 0 or 1, increase GTLREF on lanes 0 and 2.
When the failure is on core 2 or 3, increase GTLREF on lanes 1 and 3.

For dual cores: See above note in red.

Continue this process in 5 or 10 mV increments until you become "stable" on all four cores.

Record these numbers

Now continue to increase the values 1 increment at a time. Keep going until the cores start failing again and record the value at which they fail.

Then set the GTLREF for each lane HALF WAY between when they became stable and when they started to fail again, ROUNDING UP.

The low number and the high number is the top and bottom limit of the noise margin and you want to be right about in the middle. The noise margin for the NOMINAL RATIO changes with relation to FSB speed as shown in this chart: 
 
 


 
        This is noise margin tolerance, or how far away from nominal you can be with out becoming totally unstable. This window is a little wider than the nominal noise margin, but if you are in the middle of it you should be in or near the middle of the noise margin, it's a good reference point, not exact.
You can see that some Speeds have a much narrower margin than others, this can account for "FSB holes" where the computer cannot run. The flat line represents the NOMINAL GTLREF RATIO, which can actually rise or fall based on FSB speed, VTT and Vcore, so to map that we'd have to go to a 3d graph that I'm not going to make. At least with this chart you can see how narrow the margin you are shooting for is. In narrow points of this graph voltage drop due to increased impedance with increased VTT can have a significant effect on stability. That is why the increments of adjustment are so small (5mV). The "TINY" voltage drop of the lanes (typically 50mV at 1.2V) is nothing to laugh at when your window of stability is only plus or minus10mV.
Reference voltages on all four lanes are identical; it is only due to voltage drop, which can vary lane to lane, that causes settings between all four lanes to vary slightly. Even though the setting will vary, ideally actual measured voltage would be the same. The SIGNALS each lane filters are also theoretically identical wave forms so they should require equal actual voltage. Only an oscilloscope would tell you if the nominal ratio for different signals should vary, but this testing method compensates for that.
 
Since the address bus lanes are usually less affected by voltage drop you can experiment with slightly less mV on those lanes to achieve higher stability. Likely between 5 to 20 mV. You can even redo the whole method with those two lanes only once. Lane 0 and 1 are set to find
the lower value IE sets lanes 2 and 3 to 0 or 50mV so the cores fail and then following the normal testing method. Obviously for those lanes only as you are still using your original setting on lane 0 and 1. I actually believe the best results would be had using this method
and I am thinking of revising the write up to include this as part of the normal method rather than an optional step. This is especially important for dual cores, who seem to have a more touchy address bus. Theoretically, if you did it with this method, correctly, and you now measured them with a voltmeter, they would be extremely close to all, being equal in actual voltage
 
Can a wrong setting break my computer?
 
At no VTT can the maximum adjustment, +160mV, put you at the maximum 1.2 V defined by Intel for the GTLREF lanes. IMHO, Worst-case scenario a bad GTLREF setting causes all sorts of data corruption, the computer wont post and will require a CMOS reset. I do not think you can do any real damage to your CPU considering you can not exceeded, or really get too close to the maximum voltage for GTLREF lanes. This isn't to say you couldn't damage your motherboard. But I'm not going to try and test the limits because I like my hardware in a working state. Always proceed with caution when increasing voltages in your BIOS.
 
But what about the (auto) setting in the bios?
Frankly, it is more or less useless. The auto setting code is broken, as it only sets one lane (lane 1) to an arbitrary amount taken from a chart. It seems to be a vestige from when nForce boards only allowed you a global GTLREF adjustment to all four lanes.
The chart is as follows:
For a given Front Side Bus QDR (Quad Data Rate, fsbx4) speed, GTLREF Lane 1 =
0- ~1149 MHZ = 0mV
1150 - 1199 MHZ = 70mV The motherboard adds 50mV to compensate for increased impedance due to over volting and 20mV to compensate for increased freq
1200 - ~1499 mhz = 90 mV, an incremental increase of 20mV
~1500 - ~1899 = 110 mV - note that @ 1700 and up ~100mV over NOMINAL would be unstable. If VTT*.67+110mv is nominal then you'd be fine.
1900+ = 130mV
This chart is approximate because exact values of FSB speed could not be obtained due to FSB holes. I'm sure some engineer somewhere had a reason why they created this chart like this, but it's not based on anything concrete and was not implemented properly (it should set all four lanes to SOMETHING). Perhaps this will be fixed in a later BIOS release. Right now, these values are useless. It may or may not be a stable value for you (even if you set all lanes to it), this would purely be by chance if it is.
 
Again:

The 7 Series motherboard has NO WAY to monitor ACTUAL VTT voltage or ACTUAL GTLREF lane voltage.

The ONLY factor that makes the motherboard refer to the chart is the FSB QDR frequency. It does not make any adjustments when you change VTT, Vcore, or CPU multiplier, or any other setting for that matter. I have tested all of these variables and only a change in FSB frequency will cause the motherboard to refer to the chart. Changes in all the other variables will require an adjustment in GTLREF voltage, and since no automatic adjustment is made, we can conclude that setting can be disregarded. With out having the ability to monitor actual voltages of VTT and GTLREF the (auto) setting adjustment the motherboard makes to lane 1 are seemingly pulled out of thin air. GTLREF Nominal ratio does not scale this way (IE in steps).

There is no (automagic) mode in the bios.

Calculating GTLREF adjustments based on ratio:

If you wanted to find out what adjustment to make based on ratio increases, or reverse the equation to find out what ratio you are running at for each lane: 
 
[VTT*(.XX-.67)]*1000=GTLREF adjustment in mV; where .XX is the new ratio.

The reverse is:

[(GTLREF mV Adjustment in bios / 1000)/VTT] + .67 = simulated commanded ratio

I hope this puts a lot of debate to rest.

NOTE: There are some instances, like if you were running VTT 1.2V or 1.1V (impossible on 7 Series), where you would want to reduce GTLREF to be at a ratio lower than .67 in order to run an over-all lower Vcore and/or reduce core temps. You would test this in much the way you would test making increases, HOWEVER, with the 7-Series in order to set a value BELOW what the motherboard is trying to set you must modify your motherboard, and that is not covered in this guide. 45 nM dual cores have been known to run @ VTT .63, which would allow you to reduce VTT and Vcore and run a lot cooler, since CPU and FSB temperature are a function of the square of their voltage (a small decrease in voltage would result in a relatively large change in temperature and vise versa).  Running cooler and at a lower voltage is of coarse highly desirable and it is unfortunate that we cannot make negative adjustments like other motherboards.

**The 790i has the ability to adjust to negative numbers because it better supports 45nm cores, of -155mV to +155mV.

790i owners can adjust down with the same testing method that we use to adjust up if they are running at stock clocks or lower than stock voltage. This is desirable to achieve overall love Vcore and lower temps at stock or near stock speeds and voltages.

If you are looking to overclock your ram significantly, and you are willing to modify your motherboard, RAM GTLREF adjustments can be extremely helpful. RAM is normally VDIMM*.5, increasing the ratio or adjusting for voltage loss can make a huge difference in overclocking your ram, especially if there is significant voltage drop on that GTL lane of your motherboard.

Check back frequently as I will keep this thread up to date.

Sources: Intel, Thetechrepository, various online articles, days of my own testing and research.
Image Source: Thetechrepository 



                  
 

rjohnson11 


I spent A LONG TIME compiling and a lot of hard tedious work testing this information, DO NOT copy paste any of this into your own posts in other threads or other forums, rather, post a link to this thread. 
   
 

Hard to do when you can't keep this thread in the same place. I've been sharing a dead link because you ninja'd the orginal thread.
 
 
I've also found that trusting which core is failing in Prime95 is not always 100%. If the error happens in the L2 cache then it could fail any core. The best way to properly test the lanes is to adjust one lane at a time.