I've been working on a wholistic guide to designing a water-cooling system of late. Using a mix of real-world test data, and calculating pressure drops, I've been able to put together an analysis of the impact of tubing sizes on CPU temperatures.

The radiator and waterblocks are:

Thermochill PA120.2 with 2 x Yate-Loon fans at 12v

Swiftech Apogee GTX

Conroe C2D CPU, overclocked and under load, emitting 100W of heat

2 meters of tubing length

Loop order is pump->radiator->waterblock->pump

Using 1/2" ID tubing and 1/2" OD barbs, I determined the pressure-drop curve for the system. Using Swiftech's published test data for the Apogee GTX, and a flow-performance curve for the PA120.2, we're able to determine the pumping hydraulic power required to push various flow-rates. Using established typical ratios of hydraulic power to actual power draw and heat dump of known real-world pumps, we're able to throw into the mix the amount of pump heat dump required to push any flow rate. We first establish this independently of an actual pump (i.e. determine the theoretical best pump), and then select an actual real-world pump that best suits the theoretical target, and then using the PQ curve of that pump, determine the final flow rate of the system, and hence the correspondent final CPU temperature.

Now in a wholistic model, we're modelling not just the impact of the water-flow rate on the CPU temperature, but the impact of the total heat dump of the cooling system (CPU, pump, radiator fans) has on the room environment, which in turn raises the temperature of the air in the room, and so in turn raises the water temperature because the air-in temperature into the radiator will have warmed up. The effect is very small, but I still model it.

Global temp = 22C

Room C/W = 0.005

Fan Heat Dump = 2.0W

The proposed tubing sizes and fittings we'll be investigating are:

6.35 (1/4") ID tubing with quick-fit fittings

8mm (5/16") ID tubing over 6mmID|8mmOD barbs

8mm (5/16") ID tubing with quick-fit fittings

9.6mm (3/8") ID tubing over 7.5mmID|3/8"OD barbs

9.6mm (3/8") ID tubing with quick-fit fittings

11.1mm (7/16") ID tubing stretched over 10.5mmID|1/2"OD barbs

12.7mm (1/2") ID tubing over 10.5mmID|1/2"OD barbs

Quick-fit fittings are those similar to those found on the Swiftech MCW50 (http://www.swiftech.com/products/mcw50.asp)

Running the above range of tubing/fitting sizes through the optimal pump power estimator software I wrote, it predicts that the best pump to use is one that's consuming around 10-13W, with optimal pumping efficiency in the ranges of 3-6LPM. I won't go into the intricacies of the pump power estimator. It's not an exact science, suffice to say that it looks at the wholistic scenario given a waterblock, heatload, room C/W, radiator, system restriction, and so on, and puts out a suggestion for where the optimal range of pumping power lies for that setup. This allows us to then pick a real pump that closely matches the suggested pumping characteristics.

Using the Laing data here: http://www.laing.de/file/66 we see that an unmodified DDC1+ (more commonly referred to in forums as the DDC2) is a very good pump fit for our scenario. Another excellent alternative would be the DDC1 with a modded top.

Okay, so our optimised system consists of:

Laing DDC1+ (unmodified)

Thermochill PA120.2 with 2 x Yate-Loon fans at 12v

Conroe C2D CPU, overclocked and under load, emitting 100W of heat

2 meters of tubing length

For the various tubing/fitting sizes, the PQ curves for a full system for each tubing type looks like this:

I overlaid the curves onto the PQ graph for the Laing DDC1+

The flow performance curves for the radiator and waterblock are illustrated on the following graphs:

...and...

The total CPU heat load is 100W. The total system heat load is 114W . We assume a fixed 14W heat dump from pump which was derived from other testing. This does in fact vary a little as we can see by the Laing graph. As flow rates decrease, so does power draw, and therefore the heat-dump as well. For simplicity we'll assume a fixed 14W heat dump for now.

The intersections all are:

6.35mm quick fit = 4.45LPM flow, 0.0795 block c/w, 0.0374 rad c/w

8mm barbed = 4.75LPM, 0.0783 block c/w, 0.0373 rad c/w

8mm quick fit = 5.6LPM, 0.0770 block c/w, 0.0369 rad c/w

9.6mm barbed = 5.7LPM, 0.0768 block c/w, 0.0369 rad c/w

9.6mm quick fit = 6.2LPM, 0.0762 block c/w, 0.0367 rad c/w

11.1mm barbed = 6.3LPM, 0.0761 block c/w, 0.0367 rad c/w

12.7mm barbed = 6.35LPM, 0.0760 block c/w, 0.0366 rad c/w

Final CPU temperature is ambient (22C) + system load (114W) * radiator C/W + CPU Load (100W) * block C/W

The final CPU temperatures work out to be:

6.35mm quick fit = 34.21C

8mm barbed = 34.08C

8mm quick fit = 33.91C

9.6mm barbed = 33.89C

9.6mm quick fit = 33.80C

11.1mm barbed = 33.79C

12.7mm barbed = 33.77C

So there we have it. The differences between varying tubing sizes.

Okay, the more astute of you will point out that the block C/W is really the case-to-block C/W, and that the actual CPU-die-to-block C/W is a lot higher. Even if we triple block the C/W (which would be an absolute upper limit based upon older research), we get:

6.35mm quick fit = 50.11C

8mm barbed = 49.74C

8mm quick fit = 49.31

9.6mm barbed = 49.25C

9.6mm quick fit = 49.04C

11.1mm barbed = 49.01C

12.7mm barbed = 49.00C

I'll leave it to everyone's own personal value based judgement to determine the relative importance of the differences seen....

It's certainly not the 5C figure that people bandy about. I never expected that it ever would be myself. In my own testing with arbitrarily choking the flow-rate in a test-system, I've always been amazed at the low flow resilience of many setups. Below 2LPM is where things start getting pear shaped quickly for most systems. My recommendation is that even if you're a low-flow fanatic, always ensure that your flow-rates are above 2LPM at the very least, and preferably above 3LPM if at all possible. Still, even when given 1/4" tubing installed with quick-fits and a decent pump like a DDC2, we can see that flow-rates in excess of 4LPM aren't a problem.