Complete BIOS Tuning Guide - "Over Voltage" (Cont'd)

CPU GTL_REF(0~4) Ratio - [Auto, Default, -226mV ~ +224mV] Gunning Transceiver Logic (GTL) voltages are nothing more than reference thresholds used by the CPU and SPP (MCH) in determining if the voltage potential on a FSB data or address signal line is intended to represent either a logic level 1 (high) or logical level 0 (low). These voltages are generated by the motherboard, usually though the use of precision voltage dividers or high-accuracy, programmable digital potentiometers. Although the most common way to specify these values is as a percentage (or ratio) of full scale CPU VTT Voltage, NVIDIA has decided to provide these settings as simple offsets meaning a little math will be involved if you want to make a change.

These reference values are particularly important when overclocking quad-core CPUs, especially when venturing above about 475MHz FSB. The ability to tune these values per die can mean the difference between achieving a stable overclock at 475FSB and one at 500FSB. The real power seen in GTL reference tuning comes in the ability to provide each quad-core die with separate reference values. (Recall that current quad-cores from Intel are an MCM design consisting of two dual-core dies on a single substrate package.)

This can get tricky though as each die receives two GTL reference voltages - one for the FSB Data Bus (running at the full FSB transfer rate) and one for the FSB Address Bus (running at half the FSB transfer rate). This means CPU GTL_REF0 Ratio is for the Core 0/1 Data Bus, CPU GTL_REF1 Ratio is for the Core 2/3 Data Bus, CPU GTL_REF2 Ratio is for the Core 0/1 Address Bus, and CPU GTL_REF3 Ratio is for the Core2/3 Address Bus. CPU GTL_REF2 Ratio and CPU GTL_REF3 Ratio will have no affect with a dual-core CPU installed, in which case CPU GTL_REF0 Ratio is for the Core 0/1 Data Bus and CPU GTL_REF1 Ratio is for the Core 0/1 Address Bus. As a rule, adjustments are normally made to FSB Data Bus values first. These lines are consistently heavily loaded and as such are more susceptible to the detrimental effects of reduced signaling margins.

Intel's GTL specification dictates that each reference voltage should be tightly controlled as 67% of the current VTT voltage (0.67 times CPU VTT Voltage). Typically, changes in VTT are automatically compensated with standard voltage divider circuits dedicated to the generation of each GTL reference voltage. Historical data has shown that dual-core CPUs (in particular 45nm dual-cores) often clock to higher FSB levels when the GTL reference values are closer to 63 ~ 64% of VTT while quad-core CPUs usually need the full strength value or even a little more voltage (67 ~ 70% of VTT).

As an exercise in the application of basic math, consider the following. With VTT at 1.10V, the GTL reference voltages at 67% would be 0.737V (0.67 times 1.10V). If VTT increased to 1.26V, the new GTL reference voltage would be 0.844V, calculated the same as before. If we wanted to adjust this down to 63%, we would simply multiply the current VTT voltage by the difference between the actual and desired percentages of full scale VTT (67% minus 63% equals 4% or 0.04) and then convert this to mV by multiplying by 1000. At 1.26V, the result would be 50mV and since we want to reduce the GTL reference voltage we would pick the value closest to -50mV. Sometimes tweaking VTT slightly up or down can help if none of the available values is close enough the desired offset. Of course, you will need to repeat the calculation using the new VTT.

Proper GTL reference level tuning can sometimes allow for a dramatic reduction in VTT and in some severe cases even permits the use of a lower core supply voltage (Vcore). The most recent x64 build of Prime95, version 25.6 build 6, provides an excellent means for testing the effectiveness of CPU-specific GTL tuning. Simply load the program, select the Blend type Torture Test, and watch for any instance to drop out. Note the instance number (which will correlate to the core in question), reset and enter the BIOS, make a small adjustment to the corresponding GTL reference value(s), and then start a new test run. If the test runs for a longer period of time then you can be confident that you are at least moving in the right direction. Continue this process until the system is sufficiently stable with the current settings.

Those that want to know more about GTL reference voltages and related tuning issues are encouraged to carefully examine more in-depth articles on the subject posted here and here.

NB GTL_REF Ratio - [Auto, Default, -126mV ~ +160mV] The NB (SPP) GTL reference voltage provides the same functionality as those for the CPU, the only difference being that the Northbridge uses only a single value. Tuning in the right GTL reference voltage (usually near the nominal 67% value) can sometimes lead to lower stable VDIMM requirements.

DDR3 CHA(1~2)/CHB(1~2) Ref Voltage - [Auto, Default, -126mV ~ +160mV] These voltages work like GTL references for memory, except that VTT for the memory is defined as one-half VDIMM. Small adjustments in either direction can sometimes bring additional stability or higher memory overclocks; however, testing the effectiveness of any change is exceedingly difficult and time consuming. Most of the time these are better off simply left on Auto.

Complete BIOS Tuning Guide - "Over Voltage" Complete BIOS Tuning Guide - "CPU Configuration"
Comments Locked

23 Comments

View All Comments

  • takumsawsherman - Saturday, April 12, 2008 - link

    But for $400, you only get Firewire 400. Is that like a key, or something? If we pay $800 for a board, will they finally feel as though they can afford to add Firewire800, as Gigabyte did on their $200 boards like 3 or 4 years ago?

    When they talk about adding firewire itself to a board, does it never occur to them that a faster variation has existed for 5 or 6 years now? How insulting.
  • Grandpa - Saturday, April 12, 2008 - link

    It doesn't matter what the price, performance, make, or model. If the board is unstable I don't want it! I had an Abit board once with a VIA chipset. It corrupted data when large files were transferred between drives. Several BIOS updates later, with the performance down to a crawl, it still corrupted data. Because of that ugly bad memory, stability is number one important for me. So this review is very relevant to others like myself.
  • Super Nade - Friday, April 11, 2008 - link

    As far as I know, the capacitors you mention are made by Fujitsu's Media division (FP-Cap series), not Fairchild semiconductor. Fujitsu did try to gobble up Fairchild in the 80's, but the US government killed the deal. Apart from this, I am not aware of any connection between these two companies.

    Here is the link--> http://jp.fujitsu.com/group/fmd/en/services/capaci...">http://jp.fujitsu.com/group/fmd/en/services/capaci...

    S-N
  • Stele - Saturday, April 12, 2008 - link

    Super Nade's right. The vendor marking on the capacitors - which have been the same for almost all such solid electrolytic polymer caps used on Asus boards for some time now - is very much that of Fujitsu: a letter 'F' in Courier-esque font between two horizontal lines.

    Interestingly - and confusingly - however, once upon a time this logo was indeed that of Fairchild Semiconductor... the deal that almost happened in the 80s may have something to do with Fujitsu's current use of the said logo. Either way, Faichild Semi have long since changed to their current logo (a stylised italic 'f') so today, any current/new electronic/semiconductor component carrying the F-between-bars logo is almost certainly a Fujitsu product.
  • jojo29 - Friday, April 11, 2008 - link

    Just wondering how the Anandtech's Choice P5E3 Premium ( which i plan on buying) stacks up against this Striker? Any comments? Or did i miss something in the aricle as i was only able to skim through it, as im at work atm, and dontcoughwantcoughtogetcaughtbymybosscough...
  • kjboughton - Friday, April 11, 2008 - link

    We used one X48 motherboard in this review and it was the ASUS P5E3 Premium. Enjoy the full read when you make it home. ;)
  • ImmortalZ - Friday, April 11, 2008 - link

    You mention that overclocking the PCI-E bus provided tangible performance benefits on the EVGA board.

    Did you read about the rumblings around the net about some G92 based cards overclocking their GPU with the PCI-E bus? There are supposedly two clock sources for these type of cards - one on board and the other slaved to the PCI-E bus.

    Are you sure that the performance improvement is not because of this anomaly?
  • CrystalBay - Friday, April 11, 2008 - link

    Hi Kris, while UT3 does scale very well with multi-core. The game it self has no DX10 support as of yet. Hopefully EPIC will will enable it in a future update...
  • Glenn - Friday, April 11, 2008 - link

    All the benchies and comparisons are great, but how does it compare to a P35 board? A 965 or X38 board? I doubt you will convert those that already own an X48 and I (P35) have no point of reference within this article to see if I'm 5, 10 or 25% behind the preformance curve?
  • Rolphus - Friday, April 11, 2008 - link

    Interesting review... only one question though. Why use the 32-bit version of Crysis on Vista x64? Is there an issue with the 64-bit version that I don't know about?

Log in

Don't have an account? Sign up now