Intel Quad-Core Cooling: The Best from Swiftech Square-off

[freecableguy]

Intel Quad-Core Cooling:
The Best from Swiftech Square-off

Date: 30 December, 2006
Author: Kristopher Boughton

Category: Water cooling
Manufacturer: Swiftech


Waterblock design has gone through significant change over the years. Often time advances are driven by the ever-present pressure of users demanding the utmost in quality, performance, and value: which can sometimes make choosing the right waterblock a rather daunting task. What might have been the top performer for the single core Athlon 64 and Pentium4 may quickly show shortcomings when used in a test environment that includes the dual core Smithfield or Athlon X2, or even the newly introduced Intel quad core (code named Kentsfield/Clovertown). Ironically, these same advances in design can produce some rather unexpected results when applied to newly released technologies, creating a thermal solution that may be inferior to what might have been considered an older, “dated” design. While this is unfortunate, it does reinforce the need for articles such as the following, which attempt to document and explain the what, why and how of these situations.

Intel’s new quad-core Kentsfield made a strong appearance in the overclocking scene within a relatively short time. The often attacked, dual die design has created an unanticipated side effect – the concentrated cooling effort of highly machined waterblocks such as the Swiftech Storm is targeted at an area that is a partial void, one which may in fact be the geometric center of the heat flux but not necessarily the hottest part of the core. This creates a condition in which more conventional cooling designs, ones which utilizes a larger contact area, can potentially provide much better performance. Without further ado, let’s get down to how we tested.

Testing Guidelines

• Processor (CPU) frequency was varied by altering the host frequency (FSB) in 66MHz increments between 133 and 333MHz while holding the multiplier constant at 9x. This created test frequencies at 1.2, 1.8, 2.4 and 3.0GHz.
  

• Voltage ID (VID) was set to 1.300, 1.400, 1.500, and 1.600v by BIOS at each tested CPU frequency. Full-load voltages were measured using a digital multimeter and recorded to three significant digits. The CPU voltage read point was the load side of a PWM inductor/choke in the 5-phase VCC regulating circuit. A 2-input, free-standing digital thermometer was used to measure radiator fan suction temperature (intake) and exhaust temperatures to one significant digit.
  

• Motherboard Monitor 5 (MBM5) with Core Temperature Monitor Plugin v1.1 was used to simultaneously monitor and record all four (4) core temperatures as reported by Intel DTS (digital thermal sensor). Full-load temperatures are a straight average of recorded core temperatures.
  

• A Kill-A-Watt P3 was used to measure mains power draw to the nearest watt.
  

• A multi-threaded version of Prime95 (small FFTs) was used to load all four (4) cores to 100% load for 30 minutes for collection of full-load voltage and core temperatures. This time was utilized to ensure that the system was at equilibrium at time of data collection.
  

• Since room temperature could not be precisely controlled, all temperature calculations were made with respect to ambient temperature.
  

• All system components were held constant with the exception of the waterblock being tested.

Testing Methodology and Intermediate Results

While it may be common knowledge (and common sense) that core temperature varies as a function of processor frequency, voltage and overall system thermal resistance; modeling this relationship is difficult and rather impractical as the resulting mathematical equation is entirely specific to that particular system configuration (with respect to the processor, heatsink/waterblock, radiator, fan, fan speed, tubing, pump, coolant, etc.). But we did it anyway. Logically it follows that the only way to make true, direct, relative comparisons is by creating a benching/test rig in which the waterblock/heatsink is the only component being changed.

We say relative because using that waterblock with any other system configuration is sure to produce different results. In fact, system thermal resistances derived using this testing methodology can only be used for relative performance comparisons with other waterblocks/heatsinks made utilizing this specific model. It is extremely important that readers understand that while these results are accurate and trustworthy, it is impossible to make a direct comparison to thermal solutions made using alternate models. Future plans include testing as many cooling solutions as available in order to create a large database of results.

Collecting data is time-consuming and requires methodical attention to detail. For each tested frequency, each specified voltage is selected by BIOS and data is collected at the conclusion of the 30-minute full-load period.

A graph of Voltage versus ΔT (change in temperature) is then created over each frequency tested. The relationship is clearly non-linear, meaning that as voltage changes, ΔT does not change by a simple multiple of the difference in voltage. Simply put, as voltage continues to increase, ΔT increases at an increasing rate. We then use a best-fit equation to model the data. This then allows us to correct each frequency/voltage data pair to the same voltages used in the first data set (1.2GHz). This is important as the motherboard PWM voltage droop increases as both frequency and requested supply voltage increase, which would result in the introduction of small errors in the next graph (described next).

Graph 1.A: Measured Full-Load Voltage vs. ΔT for Swiftech Apogee

Graph 1.B: Measured Full-Load Voltage vs. ΔT for Swiftech Storm (revision 2)

Once we correct the ΔT (to the same 4 common voltages) we can then plot that versus processor frequency. Unlike voltage versus ΔT, this relationship is linear. This means that as processor frequency is increased (and voltage is held constant) the ΔT developed will increase as function of frequency to the first power. Note, there is an x-intercept, which for all intents and purposes pretty much gives us the maximum full-load temperatures associated with minimum operating frequencies (an infinitesimally small step above 0Hz) at the given supply voltage. We say maximum because any results returned is at best an approximation since it requires an extreme extrapolation of the model equation. This brings up an important point in that the validity of data examined outside of our testing range can be challenged quite easily as it is based on supposition of our model, if you will, and not actual results. Don't take this to mean that you can't use the final developed model to calculate full-load ΔT's outside of the range of 1.2-3.0GHz. No, far from it; but you shouldn’t expect that this model holds true for values well outside this range (like 100 MHz or 10GHz, etc.).

Like the graph above, R-squared values (a statistical measure of how well a regression line approximates real data points) are included as an indication of how well the data collected fits our expected mathematical relationship.

Graph 2.A: Processor Frequency vs. ΔT for Swiftech Apogee

Graph 2.B: Processor Frequency vs. ΔT for Swiftech Storm (revision 2)

This final graph shows us how ΔT varies with respect to CPU power / heat generation (which is in itself a function of processor frequency and voltage). This is a great response curve because it allows us to programmatically determine the processor full-load heat generation by simply noting the full-load ΔT.

The power versus ΔT graphic is actually created using two (2) sets of data. The best-fit linear regression curve (shown by the black dashed line) is based on data calculated from the advertised thermal design power (TDP), measured default frequency/voltage ΔT, and the measured ΔT of each data pair point. The individual data points (shown by blue diamonds) are based on measured mains power draw and the PSU's advertised efficiency factor (an active power correction factor of 0.99 is neglected in the calculation). Indicated error bars are +/- 2%.

The results are nothing short of remarkable. Using two independent methods of calculation, we have managed to validate measured data and testing methodology.

As a final point, note that the slope of the dashed line is our calculated thermal resistance, given in °C/W. This will become our primary basis for comparison with other thermal solutions. Lower thermal resistance means that for a given heat load (W) there will be a lower corresponding increase in ΔT. Which, of course, is of principle concern to an overclocker.

Graph 3.A: Processor Power vs. ΔT for Swiftech Apogee

Graph 3.B: Processor Power vs. ΔT for Swiftech Storm (revision 2)

Final Results and Conclusion

Shown below are the system configuration details for the waterblocks being tested. The only variance was the waterblock as it was removed and the next waterblock installed directly in its place. Deviation for loading pressure and thermal interface material (TIM) was minimized using a static pressure load system and a 'burn-in' equilibrium time of 4 hours before data collection began.

Copies of the Excel spreadsheets used to create the graphs above and the figures below, can be downloaded from the attached files field of this article, located at the bottom of the page. These spreadsheets show all mathematical models and formulas used in the development of these results.

The 'Results Calculator' portion of the XLS files (shown below in Figure 2.A and 2.B) are particularly useful as they allow the user to input desired processor frequency and voltage, ambient temperature/water block supply temperature, and determined thermal resistance and calculate the following: Full-load ΔT, Full-load temperature (average of all cores), thermal power, percent overclock, percent power increase, Thermalmarks (a new, experimental basis for cooling solution comparisons).

Figure 1: Price to performance ratio - lower is better

Based on our calculations the Swiftech Apogee is 0.015 °C/W better than the Swiftech Storm (revision 2) at keeping the new Intel quad-core series of processors cool! This equates to about a 2°C difference in average full-load temperature between these two blocks when dealing with a stock Intel QX6700 processor @ 2.66GHz. Although this difference may seem small, it is important to note that differences in temperatures of this magnitude are significant, especially when taking into account the large disparity in price. Figure 1 shows the relative cost per unit cooling potential - lower is better.

With such a large difference in both price and performance, both in favor of the Swiftech Apogee, it is hard not to give the waterblock our highest recommendation when cooling the new Intel quad-core processor. Those that are looking to upgrade to this CPU can be confident that their dollar is well-spent when choosing the Apogee.

Figure 2.A: Final Results for Swiftech Apogee

Figure 2.B: Final Results for Swiftech Storm (revision 2)

Copies of the official data collection/calculation spreadsheets and usage instructions can be downloaded below. Microsoft Excel is required to view and manipulate the XLS files. This work is Copyright © 2006 The Tech Repository (All Rights Reserved) and no portion may be used for commercial or marketing purposes without the express written consent of The Tech Repository. Distribution and licensing rights available upon request.

Questions, comments, and request for review of commercial samples should be directed to the author of this article via email.