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

[rob2k]

Just read the whole thread start to finish - well done Cryptik for such an informative topic.

For me I have noticed some conflicting posts and would like some confirmation. Also, some of the terminology can be confusing.

With a 45nm CPU the GTLs are negative so when talking higher and lower values then mathematically that means closer to zero or positive values for higher and for lower values a more negative value is needed ??!!.

If something is controlled by a formula then it beats me as to why the bios cannot calculate this. I have a 45nm CPU so my GTLs should be negative which is partly backed up by my bios having increments of 5 in the negative range and increments of 10 in the positive range. Is that intentional or just luck ?.

I use a VTT of 1.4v so my CPU GTL should be -45mv.

So where I have read that NB GTLs are generally higher than CPU GTLs I need to look towards zero or a positive number.

Could someone please clarify on a quad core which GTLs correspond to die 0 and die 1 ?. One post said GTL 0 & 1 for die 0 and another GTL 0 & 2 for die 0. I also read a post that said GTLs 0 & 2 are generally higher than 1 & 3, if this is the case is the calculated value of -45mv the highest or lowest value ?.

For example :-
GTL0 -35mv (die 0?) - higher than calculated = smaller negative
GTL1 -45mv (die 1?) - calculated value
GTL2 -35mv (die 0?) - higher than calculated = smaller negative
GTL3 -45mv (die 1?) - calculated value
(please correct where appropriate)

Sorry for such a long post but thanks in advance for any help.

[IanB]

With a 45nm CPU the GTLs are negative so when talking higher and lower values then mathematically that means closer to zero or positive values for higher and for lower values a more negative value is needed ??!!.

If something is controlled by a formula then it beats me as to why the bios cannot calculate this. I have a 45nm CPU so my GTLs should be negative which is partly backed up by my bios having increments of 5 in the negative range and increments of 10 in the positive range. Is that intentional or just luck ?.

I use a VTT of 1.4v so my CPU GTL should be -45mv.

I think you have misunderstood. The GTL values are ratios of VCC or other voltages. In some/most BIOSes you can set the ratio (eg. 0.67 x voltage) directly and that gives an absolute (positive) value, always. In some other BIOSes you are not able to do this, the ratio is set and cannot be changed, so you must enter an OFFSET to change the ABSOLUTE value to what is required, eg.:

0.67 (fixed) x voltage +/- offset = GTL ref voltage

The OFFSET can be positive or negative, but the final ref GTL will always be a positive value. It depends on what your mobo BIOS allows you to change and if you must enter an offset you must know the fixed ratio that is being used.

[DemonEyez]

I thought I had noticed my temps got higher when messing with it but, wasn't sure. I mean if you can and it helps then jack it up but, I just feel safer with low ones that's all. I could probably get my 4Ghz mark with jacking mine up but, it wouldn't be 24/7 so it. I will leave mine where it is as I am higher then most people any way with my chip where it is.

For 4Ghz or a low fsb you won't need a high vTT and can keep it a default. This only come into play with say 4.2Ghz and up of a FSB of 500MHz and up

vtt shouldn't really raise temps all that much, since its not actually applied directly to the cpu itself like vcc. vss/vtt are termination voltages and as such would at most add a little more heat, but nothing near adding an extra 0.05v of vcc would do. the only thing it would heat up would be the voltage regulation mosfets!

I noticed (slightly) higher temps when upping vTT . Going from 1.30v to 1.35v gave me a 3 degree rise, so that's not much. MosFets are cooled by heatsinks and CPU is cooled by waterblock and are not in contact with one and other

[rob2k]

I think you have misunderstood. The GTL values are ratios of VCC or other voltages. In some/most BIOSes you can set the ratio (eg. 0.67 x voltage) directly and that gives an absolute (positive) value, always. In some other BIOSes you are not able to do this, the ratio is set and cannot be changed, so you must enter an OFFSET to change the ABSOLUTE value to what is required, eg.:

0.67 (fixed) x voltage +/- offset = GTL ref voltage

The OFFSET can be positive or negative, but the final ref GTL will always be a positive value. It depends on what your mobo BIOS allows you to change and if you must enter an offset you must know the fixed ratio that is being used.

Hi Ian,

I have used the formula from post 1 Vtt x 0.667 + Y = Vtt x 0.635 where Y = the GTL ref adjustment for a GTL ref V of between 0.8 and 1.0, my bios asked for GTL ref V in milli volts rather than asking for the multiplier so for a Vtt of 1.4 Y= -45mv. I think my bios is asking for the value of Y.

[truehighroller]

For 4Ghz or a low fsb you won't need a high vTT and can keep it a default. This only come into play with say 4.2Ghz and up of a FSB of 500MHz and up

I think it might help me at that point though because I can't get my chip to go over 490 with this board. I don't know if it is just the MB holding it back or what because the chip does so well where it is at, I don't see why it wouldn't be able to run over 490 easy besides maybe the MB is holding it back or my voltages need to start being tweaked more at those high fsb settings.. That's my logic any way maybe I am wrong..

[mikeyakame]

I think it might help me at that point though because I can't get my chip to go over 490 with this board. I don't know if it is just the MB holding it back or what because the chip does so well where it is at, I don't see why it wouldn't be able to run over 490 easy besides maybe the MB is holding it back or my voltages need to start being tweaked more at those high fsb settings.. That's my logic any way maybe I am wrong..

it's not exactly the raise in vtt that increases your temps, rather I have a feeling it's more to do with the gtl reference voltage in relation to it. if you raise vtt, and leave the ref as is, it's very possible to reduce the efficiency of each clock and make the cpu work a little bit harder to do the same amount of work. i would have never even considered that something as small as this could do it a while back, but my results from all my testing pretty much conclude that this is probably the case.

[CryptiK]

Just read the whole thread start to finish - well done Cryptik for such an informative topic.

For me I have noticed some conflicting posts and would like some confirmation. Also, some of the terminology can be confusing.

With a 45nm CPU the GTLs are negative so when talking higher and lower values then mathematically that means closer to zero or positive values for higher and for lower values a more negative value is needed ??!!.

If something is controlled by a formula then it beats me as to why the bios cannot calculate this. I have a 45nm CPU so my GTLs should be negative which is partly backed up by my bios having increments of 5 in the negative range and increments of 10 in the positive range. Is that intentional or just luck ?.

I use a VTT of 1.4v so my CPU GTL should be -45mv.

So where I have read that NB GTLs are generally higher than CPU GTLs I need to look towards zero or a positive number.

Could someone please clarify on a quad core which GTLs correspond to die 0 and die 1 ?. One post said GTL 0 & 1 for die 0 and another GTL 0 & 2 for die 0. I also read a post that said GTLs 0 & 2 are generally higher than 1 & 3, if this is the case is the calculated value of -45mv the highest or lowest value ?.

For example :-
GTL0 -35mv (die 0?) - higher than calculated = smaller negative
GTL1 -45mv (die 1?) - calculated value
GTL2 -35mv (die 0?) - higher than calculated = smaller negative
GTL3 -45mv (die 1?) - calculated value
(please correct where appropriate)

Sorry for such a long post but thanks in advance for any help.

This is the thread for long posts GTL 0 is for the address clock strobe input reference voltage for die 0, GTL 1 is for the data clock strobe input reference voltage for die 0, GTL 2 is for the address clock strobe input reference voltage for die 1, and GTL 3 is for the data clock strobe input reference voltage for die 1.

Hi Ian,

I have used the formula from post 1 Vtt x 0.667 + Y = Vtt x 0.635 where Y = the GTL ref adjustment for a GTL ref V of between 0.8 and 1.0, my bios asked for GTL ref V in milli volts rather than asking for the multiplier so for a Vtt of 1.4 Y= -45mv. I think my bios is asking for the value of Y.

To get all your Vref's the same with the Vtt of 1.4v, you should for example, leave your 0.635x offset value on 'auto' and use -45mv as the Y value for the 0.667x.

So:

0.635 x 1.4 = 0.889v
0.667 x 1.4 = 0.9338v (you need to reduce this to be as close as possible to 0.889v)
0.9338 - 0.889 = 0.045
0.9338 - 0.045 = 0.8888v

Therefore use -45mv for the 0.667x GTL ref and you get the resulting Vrefs as close as possible.

[rob2k]

Hi CryptiK,

Thanks for the reply. I was trying fifferent values last night when running HyperPi but had failures on cores 1 & 2 when counting 0-3. I will have more time tonight so will disable die 1 and aim for stability on a die by die basis. I have heard people using prime and then determining whether to raise or lower the GTL ref by the amount that it fails - any more info on this or a similar method would be greatfully recieved.

It sounds like I need to find the 0.635 offset value in my bios and also which GTLs refer to the 0.667 multiplier.

I have read the article at thetechrepository and their diagram of which GTL does what differs from your explanation:-


I think this is fairly important as I understand that values for data bus are generally higher than those of the address bus.

[CryptiK]

lol I shouldn't post when I'm drunk, I got duals and quads mixed up :P

Without looking back over the papers to confirm it, that jpeg looks to be correct. However knowing exactly what each one does is of minimal importance, just finding the sweetspot for each one for whatever FSB and Vtt you are running is the goal.

[davidk21770]

Hi CryptiK,

I have heard people using prime and then determining whether to raise or lower the GTL ref by the amount that it fails - any more info on this or a similar method would be greatfully recieved.

It's more a way to tune the GTLs once you're in range. What I've read and seems to work is to step the GTLs (one at a time) to get to either side of the about the same failure in prime (or LinPak or superpi, etc.). Then split the difference. If that doesn't work, more than a GTL change is needed.

[CryptiK]

Well what I did was leave the GTL Refs on their standard settings, and set a clock I knew failed Orthos in less than 30 seconds. I began adjusting the GTL Refs, until I found settings that increased the time in which it failed. I managed to get the fail time up to 1 hour +/- a couple of mins.

Then I dropped the clock back to my previous stable clock, and was actually able to drop the vcore 1 notch and remain stable due to the GTL adjustments.

[jas420221]

First, great thread!!!

But second, and Im likely the one all wrong here... but a couple of things.

1. Where did you get your infor that GTLref is .63 for 45nm??? Generically speaking, and according to the techrepository link, its 2/3 (.667)across the board right?
2. Why, when my system is overclocked, would I still be looking for the same GTLref value as with stock? The over/undershoot voltages will vary significantly with the overclock correct?

[CryptiK]

When i first got my P45 board a few months ago, I knew nothing about GTL Refs, but I wanted to know how to set them. I found there was scattered, sparse and often conflicting information on how to set them up. I derived the 0.63x for 45nm from reading everything I could and noting what people were having most success with, and for the most part I kept seeing people saying that 0.65x - 0.70x was less stable for them than ~0.63x (this was mainly people with 45nm quads) and people with 65nm quads were able to clock better on ~0.667x. I experienced the same thing with my 45nm dual, if I set 0.667x it became unstable, the best setting was ~0.635x when using 1.2 Vtt for me.

It's just a starting point, and I made the guide to consolidate the information I'd read and my own findings together to give people a basic idea of what to set to begin with. I also explained how to set up the GTL Refs using the formula Asus use in some of their bioses as it seemed to be confusing some people. Thankfully it has turned into a healthy discussion with many people contributing.

Essentially GTL's need to be adjusted when you're trying to stabilize your cpu at higher (400 - 450+) FSB and higher clock speeds, under those levels it will most likely have a negligible effect. You start with say 0.635x if you have 45nm, and 0.667x if you have a 65nm, and tweak from there based on your FSB, clock, and Vtt, guided by the individual response of your cpu to the adjustments you make. Even if you have a 45nm you may end up at ~0.667x giving you the best results, or you may have a 65nm and end up at ~0.635x, there's no hard and fast rules to it. Generally the higher Vtt you set, the lower your GTL's need to be to compensate.

[jas420221]

Generally the higher Vtt you set, the lower your GTL's need to be to compensate.

From what I understood (and again, its likely me thats wrong here, but humor me a second)...

ASSUMPTION - when you raise your FSB and such the undershoot and overshoot levels increase with the overclock.

If that assumption is true, why would your goal be to end up with the same GTLref value you started with? From what I understood of the first post, I believe you aresaying the goal is to be 'somewhere' around there... right?

Also, on my board, one set of cores is .63 and the other is .67... any thoughts on that?

[davidk21770]

This thread is all about how GTLs work, knowing what you've set the GTLs for, and some good info on how to set the voltages around them. But since we're pushing the CPUs well beyond they're design ratings, this is just a starting place. Like the other voltages that we're upping, the GTLs won't be where they're "supposed" to be either. That's why I've been reading about how to tune them. What I'm finding is that 0.67 or 0.635 is not where I end up at any particular clock. Since I haven't stopped to fine tune any settings yet (when I break 4G, I'll do that before moving on), I may find that I end up there, but for lower OCs I was down around 0.635 and now up at 442 FSB, I find that I'm up at 0.68 (0.69 at 430). And NB's down at 0.62, but this memory and this MB don't seem to get along well -- the latest BIOS seems to be helping.

[rob2k]

Thanks guys.

Still playing and hopefully getting closer but I still have one core giving rounding errors half way through HyperPi (not alwats the same core!). I go back and change some GTLs and then the other core errors.

I have given up on the science and the .63x and .67x multipliers as I don't know which GTL this applies to in my bios. I tend to agree that at limits of hardware then the exact formula does not apply and a lot of trial and error is needed.

So I need a more pragmatic approach, I like the suggestion of isolating a core at a time. I was running HyperPi on all 4 cores but now wish to run 2 instances of it - on one core each - so that I can isolate a die at a time. The first runs will be on cores 0 & 1 = die 0 = GTLs 0 & 2 and then I'll work on die 1. I have seen in task manager using 'set affinity' how to select CPU but have do I select which core each instance (process) of HyperPi to use ?.

Just to test the positive/negative part of the equation I ran GTLs at +50 and had an average core temp of 58, at GTLs of -50 it was an average of 6 degrees lower so neagatives win for me !!.

[rob2k]

I have seen in task manager using 'set affinity' how to select CPU but have do I select which core each instance (process) of HyperPi to use ?.

Not tried this yet but I plan to create 4 shortcuts for HyperPi named HyperPi_core0 -> HyperPi_core3. The target for HyperPi_core0 will be modified to C:\windows\system32\hyperpi.exe /C start /affinity 1 and silmilarly for the other shortcuts so that they automatically run on the core I want.

[jas420221]

From what I understood (and again, its likely me thats wrong here, but humor me a second)...

ASSUMPTION - when you raise your FSB and such the undershoot and overshoot levels increase with the overclock.

If that assumption is true, why would your goal be to end up with the same GTLref value you started with? From what I understood of the first post, I believe you aresaying the goal is to be 'somewhere' around there... right?

Also, on my board, one set of cores is .63 and the other is .67... any thoughts on that?

Cryptik?

[CryptiK]

Overclocking by increasing voltages and speeds obviously alters the optimization of the circuits (which are tuned for stock speeds). Increasing voltage changes the impedance of the circuit and the amount of noise that is present, such that the more voltage that is added, the more noise that is generated. The result of this is an increase of overshoot, ringback and settling time. To try and compensate for this, we can adjust Vtt and the GTL Refs, while using the minimum possible vcore to keep noise as low as possible, and also choosing FSB ranges that are within a small margin (around +/- 15 MHz) of multiples of the cpu clock 33.33MHz. Mikey mentioned these settings earlier. You will find that when you keep within the small window around each optimal FSB frequency, that Vtt, and GTL's can be reduced to close to normal levels, while setting FSB frequencies outside these windows causes the need for markedly different Vtt and GTL settings, far from the stock values.

The goal isn't to end up with the same GTL Ref you started with, it can just sometimes end up that way by chance. You have to simply find GTL ref settings that keep V(IH) and V(IL) withing their respective operational parameters for the given FSB, clock, and Vtt. I just suggested that people set those values (45nm = 0.635x and 65nm = 0.667x) initially to start with, and tweak from there until the optimal values are found for whatever FSB/clock/Vtt you have set. Personally the optimal setting has been a little over 0.667x when using a Vtt of ~1.16v, and around 0.635x or less when using a Vtt of ~1.22v. Some people, especially with quads, seem to benefit when increasing the Vtt and the GTL Refs, it really just depends on your individual CPU. You will probably find the optimal settings to be between 0.61x and 0.70x.

Concerning the GTL Ref 0/2 = 0.63x and 1/3 = 0.67x on the P5Q-D, I can't really suggest a reason why this would be so, unless it was optimized for dual core operation ie: keeping the address bus GTL Ref higher than the data bus. (More advanced bioses allow more control over the GTL Refs allows each GTL Ref 0/1/2/3 to be set individually and this is more suited to quad core GTL Ref tuning.) However even though I used a dual on the P5Q-D, I found that by setting the 0/2 GTL Ref to 0.67x when using a low Vtt increased my stability.

[mikeyakame]

i concur with those results on a p5q. for 466fsb i ended up using 1.38v vtt, 0.635 gtl 0/2, 0.665 gtl 1/3, 1.34v vnb, 0.625 gtl nb, 100ps delay cpu clock skew, 300ps delay nb clock skew, 150ps advance dram skew a1, 100ps advance dram skew a2, 50ps delay dram skew b1, 100ps delay dram skew b2. these were pretty much on point with the corsair dominator rev1.2 micron 1gb sticks i set it up for. and these settings were adequate for upto 1240mhz dram which was as far as i bothered to push it for testing. also these particular settings for the board i was able to push the q6600 g0 upto 3.74ghz (466x8) with performance level 9, which was using 4:5 divider. that seemed about the limit for that divider with that high of an fsb, and considering i was able to get down to low 54ns mem latency at only 1120mhz dram I was pretty stoked since I've found none of the p45's can handle latencies much lower than that without encountering random memory errors.

also found the p5q deluxe suffers from pretty heavy pci bus cross talk at high fsb and nothing I was able to do would make it more than slightly manageable at best. inserting/removing usb devices too quickly would trigger occasional bsod's / reboots, pci bus device scans also caused random intermittent hanging, and there was a few other annoying little problems i found when pushing a p45 chipset at least from Asus' implementations.

[rge]

Can any of you adjusting GTL Ref quantify approximately how many mhz you gained at given vcore compared to stock GTL Ref settings? I would test it myself, except I have GB P35 DQ6 and a paucity of OC settings available in bios and of course GTL is not one of them. My E8600 does 4.2 prime 12hr stable 1.33cpuz, 4.3 at 1.36, 4.4 at 1.4. I realize all setups/mobo different but would be useful to know for example if some get ~100 mhz higher, etc on same vcore...to get an idea of what I am missing.

You guys are making me want to upgrade my mobo just to have some settings to play with, but wont do that since nehalem is here...love the GB board but bios is pissing me off.

[Demo]

I have a Q6600 @ 3.2Ghz (400FSB)
My board defaults the GTLREF at 63% ( so that would be 0.663 ? )
Since the initial post on this thread states that 0.667 is the default for quads, should i change it to 67% ( 67% is the max value also )

im using 1.335v bios volts for the cpu and the default of 1.2000v for CPU VTT
board is Abit IP35-E

[CryptiK]

63% is 0.630x, 67% is 0.67x or perhaps 0.667x rounded up. You could safely give that cpu 1.4v bios, even 1.42v bios, and 1.3v Vtt, and it may do 3.4GHz on that. Set the cpu so it is a little unstable (fails orthos/prime95 in a few minutes everytime) and adjust the GTL to 0.667x and see if it increases the fail time. If it helps, drop the cpu back to a known stable setting and the GTL adjustment may help you reduce the vcore 1 notch or more. You have to experiment, as I have said a few times now, and find you cpu's optimal GTL Ref setting for the FSB, clock and Vtt you are using.

[mikeyakame]

my q6600 needs around 1.3625v in the bios to do 3.26ghz, others ive seen are similar. to do 3.4 on mine I need 1.39v in bios. q6600 like a gtlref of around 0.63-0.65 with a balanced vtt. whatever vcore you set, aim to get vtt close to it, higher preferably. if you do that you'll find that 0.63-0.65x should be on the mark. 1.39v in bios for me i use 1.38-1.40v vtt at the sensor. if you aim too low with vtt you wont have enough gtlref adjustment, this is the way it works with 65nm chips.

[CryptiK]

Yes, that is good advice, with a low Vtt, GTL Ref adjustment is very knife-edge, the optimal range is tiny and very difficult to work with. It's like that with my 45nm dual core too.