To quote Stew from four years ago...

Quote Originally Posted by Cathar
Don't really need a massive flow rate to keep something cool. The problem, as always, is to do with convectional efficiency.

Water does hold a heck of a lot of heat per unit volume. 1 litre per minute (~0.3gpm) will rise in temperature by just 1.0C per 70W of heat. Could happily cool a 140W super-overclocked heat-monster CPU with just 1LPM, provided you can get the convectional rate up high enough.

This is where it all goes a little pear-shaped through. Silicon CPU dies are pretty small things. Achieving average convectional rates, via jet impingement, much above 50000W/m²K on a flat surface (cpu surfaces are flat of course) is very hard to do unless you're using pumps significantly stronger than what most people use for watercooling their computer.

Now at 50000W/m²K, with a 0.0001m² CPU (100mm²), results in a C/W of 0.20. Since we're talking direct-die here, this is pretty much what the cooling effect would be against our hypothetical 100mm² CPU. There's no metal conduction, and there's no thermal paste interface.

Let's compare that with some of the upper-end waterblocks.

The thermal paste barrier, per 100mm², with AS5 has been fairly accurately ascertained through a number of different methods over at Procooling, to have a C/W somewhere in the order 0.065 for a good mount (most estimates are coming in around 0.06-0.07, so we'll pick the middle average for now).

Now waterblocks, when looking at the convectional rate, can be stated in one of two ways. We can look at the actual convectional rate per unit of surface area that the water touches the metal, which is an incredibly complex way to model the waterblock's cooling effect, albeit the most accurate way to do so. Alternately we can utilise what is called the "effective convectional rate", which sums up the net effect of the additional surface area and all the little variances in cooling effect over that area, including the conductional costs of the metal in the block, and compacts it all down into what would be the "equivalent" convectional rate as if just acting on the area of what's being cooled (i.e. the CPU area)

By utilising the simpler "effective" cooling effect method, for a 100mm² CPU die, a block like the Storm/G4 is estimated to approach the 110000W/m²K mark, and the Storm/G5 approaches the 125000W/m²K mark, when matched with the uppish range end of water-cooling pumps that people use, but these values also include the conductional cost of the metal path as well.

At 110000W/m²K, the effective C/W for a 100mm² CPU is ~0.091C/W. For the thermal paste interface, the cost was 0.065, and these two are added together to arrive at an estimate of the total C/W for our 100mm² CPU, and it works out to ~0.156. At 125000W/m²K, it works out to around a 0.145 C/W.

Now in our direct die example, the C/W of a jet impingement device on a flat surface is 0.2, or substantially worse. In order to beat the waterblocks above we would need to achieve an average convectional effect of at least 70000W/m²K over the direct-die impingement of the CPU surface area.

This is where it gets a little difficult to construct such a device, and have it work with ordinary pumps. An jet impingement array, while good for larger areas, results in localised "dead-zones" where the jet washes meet. When impinging on a conductive plate of metal, this isn't so bad 'cos the metal just conducts the heat to where it's cooler, but when dealing with the surface of a CPU die we cannot afford this - the CPU will get MUCH hotter in these regions - which pretty much forces the single jet model for jet impingement cooling of a CPU die.

The other concern is the "wall" region of a jet impingement effect. The "wall" of a jet impingement effect occurs after about r=2.5d, where d is the diameter of the jet orifice. Once you get beyond the "wall" diameter, the convection efficiency drops off fairlu quickly. This places some important to consider restrictions on the size of the jet orifice. Ideally, in fact, the jet orifice should be tuned on a per-cpu-die basis.

For a 10x10mm CPU die, the diagonal area is ~14.14mm, which means that our jet orifice should at least be 14.14/5 = 2.83mm (7/64") in diameter, and the jet of course positioned centrally above the CPU die.

Using a variety of calculations which I won't get into here, the correspondent jet velocity for a jet of this size to achieve a net cooling effect > 65000W/m²K, is 11m/s.

Using yet further pressure drop calculations (which I also won't get into here), it all works out to requiring a pump that can deliver 4.7LPM (~1.25gpm) against a pressure resistance of around 9mH2O, and that's just for the waterblock alone. All up, we'd be talking about needing to use at least something like a US-Spec (60Hz) Iwaki MD-30RZ pump in order to deliver a cooling effect that would begin to outclass what top-end waterblocks can achieve.

If the die-size is larger then the problem becomes even harder to solve for the direct-die effect to achieve something better than what a good closed waterblock can achieve.

So the end answer is, yes, it can work IF you give it a strong enough pump (and by strong - we're talking REALLY strong), and you're blasting the bejeezus out of your small and fragile die of CPU silicon directly.

When coupled with all the other risks associated with direct-die cooling, my personal opinion on the matter is that it's simply not worth it.
Original thread: http://hardforum.com/showthread.php?t=895401

Old information, yes... but still useful/informative. Either way, direct die cooling always makes for a fun little project.