Delta Temperature Range for Overclocked Quad Q6600

[jchunter]

Reason for this test: I recently replaced my Arctic 7 Freezer Pro heatsink with a Tuniq Tower 120, in order to bring down overclocked CPU temperatures. I was disappointed with the results because I had read reviews that implied that the Tower was capable of holding temperature rises to a dozen degrees, or so. Anyway, it isn't that simple and these tests shed some light on the heat removal situation.

Procedure: Quad CPU Q6600 Overclocking temperature difference between idling and four cores 100% busy with Prime95 (Small FFTs) after 5 min with Tuniq Tower heatsink. Ambient Temp: 23 - 24 C. I used CPU-Z to measure Vcore under load and SpeedFan 4.34 to measure core temperatures.

CPU Clock
freq (GHz)_Vcore_____T-idle__T-load____Delta T (C)
2.4_______1.120_____36_____48_____ 12
2.6_______1.168_____36_____51_____ 15
2.8_______1.240_____37_____56_____ 19
3.0_______1.280_____37_____59_____ 22
3.3_______1.280_____38_____62_____ 24
3.4_______1.312_____38_____64_____ 26 Edit: added 10/30
3.488_____1.312_____38_____64_____ 26 Edit: added 11/24
3.52______1.376_____39_____71_____ 32
3.6_______1.400_____39_____78_____ 39 Edit: Added 10/31


Conclusion: The temperature difference between idle and full load conditions varies from 12 to 39 degrees C, as clock varies from 2.4GHz to 3.6GHz.

The temperature difference is determined by the power (heat) produced by the CPU package and the heat that is removed by the heatsink.

Heat generated can be approximated as fCV**2, where f = clock frequency, C=total on-chip load capacitance, and v is the Vcore voltage. Heat is directly proportional to clock rate and proportional to the square of the voltage. BTW, an E6600 duo should be generating about half as much heat as a Q6600 (because it has half the load capacitance).

In this test, both Vcore and clock frequency are varied across much of the air-cooled overclocking spectrum.

Bottom Line: We have to know both clock freq and Vcore to make any sense of whether a heatsink is working effectively, or not. My next step will be to check the heatsink and chip for flatness and see if I can improve the situation by lapping either or both...

[Spawne32]

deltaT is measured as Load average - ambient= dT.

[jchunter]

Spawn,
The ambient temperature during my test was 23 C, although the quad seems to make its own microclimate near my desk, and is probably a couple of degrees warmer.

Update: I pulled the heatsink and checked for flatness - found it slightly convex and fairly rough, so lapped it under running water, using 600/800/1000 wet/dry sandpaper on a pane of glass. I checked flatness with a box cutter blade, which is long enough to span the copper block in several directions.

The Q6600 was quite flat, already. There was a very slight bulge on one side that did not seem to be worth bothering with.

This time I tried Arctic's instructions for applying AS5 to a quad (single horizontal line across both dies). Frankly, I think it makes better sense to have two vertical lines (the size of a long rice grain), one over each die, but I'm not going to knock the recommended method until I have tried it. I'm going to let the stuff settle in for a few days before I do any serious temperature measurements.

About the Tuniq Tower review (http://www.anandtech.com/printarticle.aspx?i=2906) that I refered to in the OP, it reported a ceiling temperature of about 50 C, at full load, while overclocking an X6800 duo up to 3.9GHz. This is what gave me the expectation that the Tower could handle twice the heat from the Q6600.

As I reported on the OP, my Q6600 ran much hotter at 3.5GHz.

Besides the power difference, the study used the "Far Cry River Demo" to put "near constant 100%" load on the cores, while I used Prime 95 Small FFTs to put 100% load on all four cores. Could their phrase "near constant" be somewhat inaccurate?

Bottom line: The study reports about a 10 - 12 C delta between idle and "full load" at 3.9GHz, while I was seeing over 30 C at 3.5GHz... I will report load temps after the AS5 cures a bit longer.

[jchunter]

Quad Overclocking Power/Heat

It is sobering to realize just how much extra power is consumed when overclocking a quad. All of that power turns into heat that has to be pulled out through the top of the CPU package.

The data that I posted above can be used to compute the extra heat produced when overclocking the Q6600 from 2.4GHz (stock) to 3.5GHz. Some simple math* shows that the heat more than doubles (actually 2.2 times as much heat is produced at 3.5 GHz). For a quad, which already cranks out twice as much heat as a Duo, this is an enormous burden of 4.4 times as much heat as a stock Duo.

This explains why we cannot easily extrapolate the results of earlier heatsink tests that were done with Duo processor chips, to Quads.

However, it does not explain how a measured delta T of 12 deg (idle vs load) turns into more than 30 degrees in my quad.

The heat has to pass through the top of the CPU package, which is the same size for a Quad as a Duo, into that big copper block in the Tuniq Tower. The heat transfer mechanism has to be almost entirely conductive until it reaches the heat pipes, where the heat transfer is considerably (hundreds of times) more efficient than thru copper. This probably explains why the heat pipes feel cool while the chips are running at 70 deg C.

This package junction acts like a bottleneck that forces the temperature on the chip side higher until the heat transfer rate equals the rate of heat produced. But this should be a linear process: doubling the temperature difference across the package junction should double the rate of heat transfer. All other things being equal, I would expect to see a 24 degree difference between idle and load.

In any case, the general quad cooling problem seems to be one of diminishing returns because of the high thermal resistance at the package junction.


* Using capacitive load formula P=fCv**2, the relative power is the ratio of
Power(3.5GHz) = 3.5 * C * 1.376 * 1.376 = 6.62 * C
Power(2.4GHz) = 2.4 * C * 1.12 * 1.12 = 3.01 * C

Relative power = P(3.5GHz) / P(2.4GHz) = 6.62 / 3.01 = 2.2 times as much heat at 3.5GHz.

[jchunter]

My delta T did not improve, so I pulled the Tuniq Tower and checked the AS5 coating. This was the one where I followed the quad instructions with one horizontal line that crossed thru middle of both dies. One section on the upper right corner, extending half-way down the package had no AS5. So much for the recommended quad method. There were also a few very thin spots

I cleaned up the heatsink and the CPU and decided to lap the CPU to eliminate that slight "bump" to improve overall flatness. I covered the CPU contacts with a layer of coffee filter paper and taped the edges. I dry sanded, starting with 600, then 800 / 1000. the CPU is now very, very flat. I reapplied AS5, using two vertical long rice grains (one over each die), also repolished the TT copper block and screwed it down, once again.

Immediate results were, at most, 1-2 degrees. This has not improved over 2 days.

I think that the interface between the CPU package and the Tuniq Tower is of good quality. It might get better over time but I doubt that it can reduce the delta T to a safe enough level for the Q6600 at 3.5 GHz...

[jchunter]

I am concluding that the Tuniq Tower simply is not up to the task of cooling a heavily overclocked Quad while maintaining "safe" core temperatures. The evidence points to the thermal resistance between the top of the CPU and the heatpipes. The merely warm pipes leaving the copper block are strong evidence.

There are a number of reports that water cooled quads are able to hold a much lower load/idle delta. If these reports are reliable, then the thermal resistance may be less in the package junction and more in the Tower's thick copper heat sink. The heat has to conduct through about 1/16" of copper before reaching the heat pipes... (Anybody have access to a milling machine?)

The extra copper mass that extends way beyond the chip package top doesn't help, either, because it slows up the TT's response to changing heat loads. This can be seen when Prime 95 starts running, as the core temperatures scream straight up more than 20 degrees in less than 1 second (with 3.5GHz OC).

Bottom line: The Q6600 package junction is probably as good as it can be and still shows a 30+ degree temperature difference between idle and 3.5GHz load.

I would like to hear from anyone who has demonstrated better results with the Tuniq Tower 120 (or any other air cooled unit).

[thideras]

There are a number of reports that water cooled quads are able to hold a much lower load/idle delta. If these reports are reliable, then the thermal resistance may be less in the package junction and more in the Tower's thick copper heat sink. The heat has to conduct through about 1/16" of copper before reaching the heat pipes... (Anybody have access to a milling machine?)

Those pipes are not just copper. They have a liquid/gas in them so that when heat is generated at the processor (where the liquid/cool part) is, it heats it up (remember high pressure will move to low pressure, always) so it moves the warmer liquid/gas to the cooler heatsink. It hits the heatsink, liquefies and goes back to the processor via "channels". They are not just empty

[bossman150]

I would go back to the freezer 7 pro. I am running @ 3.4 24 hours of prime the highest temp I got was 56.

[thideras]

I would go back to the freezer 7 pro. I am running @ 3.4 24 hours of prime the highest temp I got was 56.

What is the TJunction in coretemp? If it is 85, you have to add 15c to your temps Which is WAY higher than your 56c....

[bossman150]

What is the TJunction in coretemp? If it is 85, you have to add 15c to your temps Which is WAY higher than your 56c....

Nope its 100. Speedfan reads my temps 15c lower.

[jchunter]

Bossman,
Speedfan 4.33 reads quad core temperatures 15 degrees low.
What CPU chip do you have and what are your core temps when idling?

Edit: I have a stable OC at 3.4GHz, which I added to the OP. It shows a 26 degree C difference between idling and Prime95 loaded core temps.

[bing]

OT , SpeedFan 4.34 beta 34 (26 oct 2007) -> you must regged to get this beta

This beta includes a huge amount of new code and of code rewritings. It fixes the crash on some DFI motherboards and properly identifies one revision of the IT8705F chip. There is a lot of code to improve nVidia support. Intel CORE temperatures should now properly work on all systems. Added support for the SMBus available on several new SouthBridges. Added support for new chips. Added preliminary CPU identification routines needed to better support new hardware.
Beta 19 improves CPU identification routines. Beta 21 improves identification and should detect DDR3.
Beta 23 fixes a bug in charting, improves CPU identification, adds support for EMCT03 and, by popular demand, sets Tjmax=100C for Intel Core Stepping L2.
Beta 28 greatly improves Intel CPU identification, supports more hardware (like the W83793G), fixes a bug with charts and adds, in CONFIGURE / ADVANCED, a new setting to see Intel Core temperatures in a RELATIVE or ABSOLUTE way.
Beta 29 refines several things and improves CPU identification.
Beta 30 improves CPU identification.
Beta 31 improves identification and adds support for DTS readings on all Intel CPUs that support it.
Beta 32 improves identification and adds some other minor fixes.
Beta 33 restores DTS readings even if CPUID is limited through BIOS and adds native nVidia video card (excluding 8xxx) internal temperature reporting (run with /DONVIDIATEMP command line parameter to enable it).
Beta 34 adds full support for VT8251 SMBus and greatly improves nVidia support.
In order to run this beta, simply install SpeedFan 4.33, then replace the EXE and then run SpeedFan as usual. Please, use the SEND REPORT facility on the INFO tab to help me to evaluate the improvements I coded.

[jchunter]

I added another data point at 3.6GHz to my original post that shows delta T at 39 degrees - way too high, IMO. The Vcore was overspec at 1.4v at full load. Delta T accelerates rapidly with CPU rates over 3.4GHz. It seems that the Tuniq Tower just can't hack serious QUAD overclocking.

I am going to back off my operating overclock to 3.4 GHz, which is comfortably within voltage and temperature specs.

FYI, I had been able to run comfortably at 3.3GHz with the Arctic Freezer 7 Pro, even with its Mickey-Mouse push pin mount. In retrospect, it was not worth all the effort to remove the motherboard in order to install the TT.

[jchunter]

Here is a graph that shows how full load temperature abruptly increases when the Q6600 is overclocked beyond 3.5GHz. It seems that the Q6600 / Tuniq Tower combination "hits the wall" and just can't remove heat fast enough to keep temperatures under control.

For example, I have stable operation at 3.488GHz with a safe 26 deg C delta T (load - idle) and increasing frequency to 3.52GHz (just a 32 MHz increase), causes a jump to a 32 deg delta T.

BTW, the delta T measurements are remarkably constant, regardless of ambient temperature.

This should serve as a warning to overclock in very small steps as you approach 3.5 GHz.



Edit: Here is a better chart of the actual data points that better shows how steep the thermal wall is:

[jchunter]

This chart shows the CPU power, computed using the measured frequency and Vcore (load). This shows no abrupt increase in power above 3.5GHz. Therefore, the Tuniq Tower is the probable cause of the thermal wall. It simply can't remove heat fast enough.



I would be very interested in seeing a similar chart for the Q6600 with the Thermalright 120 Ultra Extreme.

[jchunter]

Here, I have added a Delta T trendline (black), based on CPU power, that shows the overclock possibilities with a heatsink that doesn't choke (no "thermal wall"). In short, 3.8GHz should be possible, while holding T(load) - T(idle) under 30 degrees and max core temps under 65, with 22 degree ambient.



CPU Frequency is also a possible limitation but this chart shows that if you are willing to overclock 5% over max freq spec (1.440 v.), 3.8 GHz is possible.



The fundamental question is whether the Thermalright 120 Ultra Extreme also has a Thermal Wall for the Q6600. Some threads on this forum suggest that it does...

If not, is it possible to mount one on the Tuniq Tower mobo bracket??? If so, I would not have to remove the mobo to install it.

[bing]

Great observations & report, and really appreciate those long hours of testing !

I've done similiar tests my self on TT120 and TRU120E, although it was done briefly and without an intense documentation like yours , the wall of TRUE120E is higher than Tuniq. Sorry, can't remember the detail, it was written on a piece of scrap paper and its gone.

Yes, heatsink has their limitation on their effectiveness, simply put, their thermal resistance is not constant, and it depends on the temperature and the wattage they're dissipating, yeah, its a complex thing if you drill deeper.

Although I'm not the expert in these heat pipes technology, it looks like the selection of the pipes it self is a crucial for a CPU HSF performance.

Again, great job & thanks !

[thideras]

I LOVE the graphs!!!

Great job on that, I will have to do that myself!

The first one (comparing temps with work) is perfect, you know that at 3.5, you get a huge temperature increase without it doing much more work. You found your sweet spot.

[Spawne32]

What are we trying to prove here again?

[thideras]

What are we trying to prove here again?

How awesome the quad is?

Why do we need a reason to test for this?

If we needed a reason for everything, most overclocking wouldn't exist