Hey people
Just wanted to share some good vibes here!
Background story As some of you who now me a bit from past (I abandoned this forums for over a year) I was messing with GTLVREF manual settings in past to no avail. I was trying to find a sweet spot based on trial & error method which I utterly failed on (too many variables for a quad core chip) .. and wasted over a month on doing so ... well it wasn't just me, there were much more people involved, but only a few succeeded in achieveing better stability at lower FSB termination voltage (VTT) with adjusted GTLVREF, given the overclock is perfectly stable to begin with - goal is only to lower the VTT ... and possibly lower the VCC (core voltage) a little at same time too.
Recently I found an old thread (linked not so long ago by VistaHead) where tt0ne posted some interesting information:
http://forums.evga.com/tm.aspx?m=835180 here are the very informational links (general info) provided by tt0ne in above thread (no need to read all):
http://www.anandtech.com/show/2482 http://www.anandtech.com/show/2500 http://www.anandtech.com/show/2404 http://www.thetechreposit...wthread.php?t=87 So, here is the deal. It worked for me (worked differently for tt0ne though hehe).
based on above knowledgeable articles, especially the one on ASUS Striker II Extreme board and their GTLVREF adjustment example, here below I was able to lower VTT from 1.25V to 1.15V (100% stable) at 450MHz FSB:
http://www.anandtech.com/show/2500/18 Point is that assuming the 790 board per Intel reference specs automatically adjusts (at least data lanes, as addressing lanes seem to be untouched on 790) to 67% of current VTT (FSB termination voltage) and a fact (some statistical info quoted by Anandtech) that 45nm quads like running at 67-70%, while 45nm dual cores like stability at 63-64% then depending on which chip type you have, you will be going either to manually reduce the GTLVREF or increase it, based on whatever the board/BIOS shows currently at given FSB speed for data lane(s).
Believe it or not, but I succeeded in actual calculation of all 4 GTLVREF values for my 790 Ultra board and Q9450 chip giving me perfect stability at 450MHz FSB - being 1800MHz for linked/synced DDR3 RAM (3.6GHz on my chip with 8x multi) at only 1.15V VTT (FSB termination voltage)... now consider this, with GTLVREF in automatic mode the lowest VTT making this overclock stable was ... 1.25-1.30V
Calculator I created a downloadable Google Documents spreadsheet. Just input desired VTT (FSB termination voltage) at given automatic GTLVREF values (790 shows those only for data lanes, it doesn't adjust the addressing lanes) and the script will calculate the rest for you (guessing address lane value)
https://spreadsheets.goog...S&authkey=CK7OwKUF Use option File -> Download as -> Excel to get a local copy of this scripted file. Input your own numbers in yellow fields only. The results are displayed in Cyan fields. Dual Core results are untested (and likely wrong as only 3% reduction was used, while 4% might be needed for better results).
How this works The idea is to assume that automatic 790 chipset GTLVREF adjustments (at least for data lanes) put the numbers at exactly 67%, so it needs to be further positively adjusted for a quad core chip (increase actual GTL voltage by 3%) or negatively adjusted (reduce actual GTL voltage by 3%).
Addressing lanes are a guesswork, I only assumed that initially they should show same values as data lanes and then I was adjusting them (either up for a quad, or down for a dual) by HALF of adjustment given to data lanes.
I'm pretty confident, based on my own experience that data lanes adjustments are close to stable, however addressing lanes are more guesswork, so that's what I did and it worked
for me at least.
Assuming that 790 Ultra board does SOME automatic adjustment and actually shows automatically adjusted GTLVREF values for data lanes (first two GTLVREF lanes with 45nm quad, only the first GTLVREF lane with 45nm dual)
Again, it may or may not work for you, especially if you have a different variant of board or chip than I do. The results may get you closer to success, so you will know soon upon testing with Prime Blend (30 mins stable is good enough), if system freezes using desired VTT and adjusted GTLVREF then something is wrong.
Check for how long each time it takes for the test to freeze the system, this way you will know if you are adjusting them in right direction.
Very bad values will cause a system freeze in seconds, while better fitted values will be able ti run Blen test for minutes before it freezes (usually in less than 10 minutes).
Eventually once you pass 30 mins in Prime Blend x64 26.6 (both 720K FFTs and 8k smallFFTs tests) then run 10 rounds of Intel Burn Test 2.51 in Max stress mode and make sure it really is stable. It was for me
I have notes from every step of process, I even tried (again) some manual trial error once I got closer, but my problem was settings all lanes (both data and both addressing) to same value. After I adjusted addressing lanes to half of relative adjustment (calculated offset) set on data lanes, it just worked and passed all tests like a charm.
A word of note on GTLVREF addressing lanes (last one for dual and 2 last ones for quad, as shown in 790 BIOS).
Since 790 chipset board does NOT adjust or show addressing lane GTLVREF in automatic mode, then it can be assumed that initially those lanes *should* be set to same value as data lanes.
It's not a problem at lower clocks because the window of opportunity is wide enough to gain stability (at higher VTT), but surely it limits quad overclocking at higher FSB. Also dual cores may be actually much easier to overclock on 790 board, because it seems the board does a better job in automatic GTLVREF adjustments for those chips. DOn't have a dual core so can't check if calculations are anywhere correct for those.
This thinking (setting all lanes to same values as tt0ne did -35 on all lanes with his setup) got me MUCH closer to gaining some stability at 450MHz FSB (1800MHz on linked/synced DDR3 RAM) and using only 1.15V FSB (VTT), but Intel Burn Test BSOD'd (stop 124) after 4 rounds and Prime Blend was freezing after like 6 minutes.
Please do not flame this thread (that's what happened to old now archived thread on same topic) as it WON'T work for many of you, sometimes your setup will be totally different, some other times your specific hardware (chip + board combo) will yield different results.
I am curious to see if those scripted calculation work for anybody else besides me
My own examples 1. 401-450MHz FSB range a) Former stable 1800MHz/3.6GHz clock (450MHz) voltages at GTLVREF on auto (showing -50,-50,0,0 in BIOS):
vcore 1.437V (as set in BIOS, actual 1.36V) - 1 notch lower value was sometimes failing the test vtt/fsb 1.25V-1.30V (1.25V was causing this overclock to freeze sometimes when gaming) spp 1.55V (likely overshot by 1 notch, but 1.45V was giving me errors under IBT and SLI was flakey) mcp 1.50V memory 1.7V (2x2GB p1p2 enabled, 8-8-8-22-2T) Stability tested in SLI mode enabled under Intel Burn Test 2.51, 10 rounds in max stress mode. b) Current same stable voltages for same speed as above, just GTLVREF adjusted to -15,-15,-30,-30:
vcore 1.431V (as set in BIOS, actual 1.36V) vtt/fsb 1.15V <==-- THIS IS FULLY STABLE NOW spp 1.50V (1.45V is fully stable under IBT and Prime Blend, but causes problems with SLI and watching movies) mcp 1.50V memory 1.7V (2x2GB p1p2 enabled, 8-8-8-22-2T) Stability tested in SLI mode enabled under Intel Burn Test 2.51, 10 rounds in max stress mode. also 30 mins of running Prime Blend x64 26.6 (720K and 8K passed). 2. 451-500MHz (or more?) FSB range a) Former stable 1866MHz/3.73GHz clock (466MHz) voltages at
GTLVREF on auto (showing -80,-100,0,0 in BIOS):
vcore 1.491V (as set in BIOS)
vtt/fsb 1.30V spp 1.55V mcp 1.50V memory 1.8V (2x2GB p1p2 enabled, 8-8-8-22-2T)
Stability tested in SLI mode enabled under Intel Burn Test 2.3, 10 rounds in max stress mode and over 8hrs of Prime Blend. b) Current same stable voltages for same speed as above, just
GTLVREF adjusted to -40,-60,-60,-80:
vcore 1.491V vtt/fsb 1.20V <==-- THIS IS STABLE NOW spp 1.55V mcp 1.50V memory 1.65V (2x4GB p1p2 enabled, 9-11-9-27-2T)
Stability tested with 32 mins of running Prime Blend x64 26.6 (720K and 8K passed). Stress Testers used: a) Intel Burn Test 2.51:
http://majorgeeks.com/IntelBurnTest_d5987.html b) Prime95 x64 26.6:
ftp://mersenne.org/gimps/p64v266.zip
post edited by feniks - 2011/06/12 09:59:24