DFI UT P35-T2R: Tweakers Rejoice!

Quad-Core, Part 3

Finding stable VTT/GTLs @ 3.6GHz with G0 stepping quad-core CPUs

Assessing a motherboard's quad-core overclocking capabilities is best done at settings which are not at the extreme limits of attainable performance. The primary reason for this is because GTL settings become ever more sensitive to one digit (0.01v) changes as we near the maximum chipset capabilities. Our experience shows that 440-450FSB on this board is a good reference/starting point, as all chipsets should be capable of attaining this level of FSB with current G0 stepping CPUs, using a "relatively minor" (minor in terms of GTL tuning) amount of tweaking. Pushing high FSBs too soon is likely to bring mixed results, without attaining sufficient user understanding to make improvements when failures are experienced.

Although there is no adherent logic that can be applied to making GTL adjustments (without a scope), having a scale does give us the chance of some success. A safe point for starting out is around 400-440FSB using around 1.4V-1.5V VTT and appropriate GTLs. This will allow the user to get a feel for VTT/GTL change requirements as the FSB speed is scaled upwards. Below is the approach we used to get our foot in the door;

Recommended Overclocking Base Settings and Components

• Intel Quad-core CPU - G0 stepping
• A sturdy and proven PSU: we used the OCZ ProXStream 1000W for the majority of our testing
• High quality DDR2: you generally want RAM that is capable of CAS4 operation at DDR2-1100 with speeds up to DDR2-1280+ at CAS5
• High performance GPU: one that is capable of withstanding high PCI Express speeds such as the NVIDIA 8800 GTX Ultra
• Good Cooling: Thermalright Ultra-120 (with high CFM fans in push/pull) or very good water-cooling (i.e. Swiftech Apogee GTX/D-Tek Fuzion block)
• High Speed 80-120mm fans: used to cool NB/PWM and RAM
• Microsoft XP SP2 for 2GB or Vista 64-bit for 4GB Memory configurations
• Target CPU VCore is 1.37V-1.43V for 450FSB @ 3.6GHz. (Good cooling is required!)
• Target VNB is 1.55V-1.65V

An Attempt at a Logical Approach to GTL Setting Adjustments

After entering target timings and voltages in the BIOS, we begin with some Prime95 testing in order to determine the effects of applied VTT/GTL adjustments. At this point we should already have determined sensible timings for RAM, which should not induce failures while we are checking for FSB stability. Sticking below default memory timing specifications while using specification voltage is a very wise choice at this point (use enough voltage for the RAM to run its specified speeds). We probed actual VDimm levels with a DMM to ensure that BIOS set values are not adrift from user set BIOS values. We could not find any significant level of voltage droop/undervolting/overvolting under idle or load conditions.

Our recommendations in terms of RAM are that you choose a good kit that can run at least DDR2-1066 @ CAS5. (2GB or 4GB of OCZ CAS5 PC9200 RAM has proven very successful with the 913 BIOS.) This will push towards the best performance possible from this board. At this time the P35 Chipset seems to favor Micron D9GMH based modules over the highly regarded D9GKX bin. A reason for the preference of the lower bin is unknown at this time. Our D9GKX modules required higher levels of voltage than we were accustomed to providing them, and we were not able to match the D9GMH based module results in terms of speed, stability, or timings ranges. The advantage to the user is of course good performance for less money, as D9GKX based modules are typically more expensive than their lower binned counterparts.

Lower speed-binned RAM is usable by adjusting strap/divider/FSB combinations in BIOS, to stay within RAM capabilities at the expense of overall performance. Tests between FSB and chipset strap ratios show the performance penalty in real world applications and games benchmarks is very small.

Of paramount importance is a level of patience and perseverance. GTL tuning takes time but can be very rewarding in terms of performance advantages. Again, perseverance is the key! Now, this is what we did to help determine which settings to change.

1. Run Prime95 in XP or Vista and note which cores fail. The reason the Prime95 Torture Test is recommended for this situation is that we can study which cores are failing the test consistently and then attempt to set the corresponding CPU core GTL values to improve failure time frames or eliminate failures altogether.
2. Repeat the Prime95 Torture Test several times using the same settings to pick up patterns of time and core related failures (if you are experiencing them). At this point, as stated earlier, it is important not to be overzealous with tight RAM timing settings.
3. If there is no pattern (i.e.; all cores fail at random), then consider tuning NBGTL by 1-2 digits in either direction, then monitor the effects again to look for changes from previous values. VCore, VNB and VDimm are all likely candidates if not within range.
4. If GTL CORE 0/1 and GTL 2/3 values are close enough to optimum for your CPU, it is likely that tuning NBGTL will create a more predictable and consistent failure on one or two of the four cores. For most CPUs you will not have to stray too far from the NBGTL scale value to find this point, though we have heard of reports where some CPU/motherboards prefer the NBGTL at around a 66% mark of VTT.
5. 1-2% variance between GTL setting adjustments sounds remarkably close between systems in theory
6. Move the corresponding GTL scale for the failing core by adding or subtracting one digit from the recommended values in the VTT/GTL table, do not stray over 1-2 digits at a time, it is imperative to check for Prime95 failure times after EACH change in GTL value.

Looking at the scales listed earlier, we can see that if the BIOS value for GTL is changed by five digits either way it only amounts to a change of only 1% between the ratio of GTL and VTT (in most cases). A minimum of five reboots and testing for a proposed 1% shift of all three GTL settings together (let alone individually) is usually necessary. GTL adjustment needs time investment by the end-user who really wishes to push the board as far as possible. At near maximum chipset capability, any single change of one BIOS GTL digit can result in the difference between a boot and non-boot situation. Of course, this is all part of the "fun" (Ed: well, for some of us) when using DFI motherboards. In the case of GTL adjustments, the added options though seemingly complex are somewhat fruitful if driving for every last MHz from the CPU. More vigorous GTL tweaking is really only required when we are trying to attain maximum FSB speeds.

Remember up to the 913 BIOS the actual GTL core functions are mislabeled. The actual core each function relates to is shown below:

CPU GTL 1/3 REF Volt = GTL ref 0/1 (Cores 0 and 1)
CPU GTL 0/2 REF Volt = GTL ref 2/3 (Cores 2 and 3)