With the miniaturization of the internal components of processors, the cooling of the heat they cause is causing quite a lot of headaches for engineers. Currently, processors have a metal IHS that makes contact with the die of the processor, and to which we install a heatsink to maintain the good temperature of the device, but would it be possible to cool the die directly with water ?
The cooling process of a processor involves having to transfer heat through various materials: from die to thermal paste or solder, from this to IHS, then conventional thermal paste, and then the heatsink. It is proven that water cooling is more efficient than this whole process, especially if we could apply it directly to the die because we would skip a good portion of the heat transfer process, but is this possible?
Processor cooling a critical factor
When desktop processors first crossed the 1,000 MHz speed barrier, it seemed for a while that you couldn’t go any higher than that. However, progress in operating speed eventually came to a halt, not because of anything to do with speed itself but because of the power requirements and the heat it generated.
Even with now common heatsinks and fans, as well as liquid cooling systems, heat is still a limiting factor that often slows the development of current processors.
Part of the problem with liquid cooling solutions is that they are limited to having to take the heat out of the chip and into the water in the first place, which has led researchers and engineers to consider passing the liquid through the chip itself.
Refrigerating the die directly with water is a reality
And this is precisely what a group of Swedish researchers have done: they have managed to design a chip and its cooling system in a single unit , with liquid channels in the chip itself placed together with the parts of it that generate the most heat, with more than promising results in terms of thermal performance.
Part of the problem with removing heat from a chip is that it generally involves multiple connections – from the chip to its encapsulation, and from the chip to the heat sink. While measures can be taken to improve these connections (the thermal paste that we mentioned before) they have proven to be inefficient, which adds to limit the heat that we can extract from the chip.
This is similar to today’s liquid cooling systems, which use liquid coolant to replace the metal heat sink. While it might be possible to place the chip directly in a heat-conducting liquid, that liquid in turn must be insulating and not undergo chemical reactions with electronic components, both obstacles that water cannot clear.
There have been a number of on-chip liquid cooling demos. This is generally a system in which a device with a set of liquid channels is fused within the die of the chip itself and a system pumps the fluid through it. This can cause the chip to heat up, and initial implementations have found that there is a small trade-off: it takes more power to pump water through these channels than is consumed by the processor. That energy is not used where heat is a problem, so it does not interfere with heat dissipation, but it does reduce the energy efficiency of the system.
In other words, they have managed to create a chip with nanotubes with coolant that significantly improves its temperature, but the disadvantage is that this requires a pump that in turn generates heat and consumes energy, so the system is not efficient. If you separate the pump from the chip, the galor it generates is not a problem, yes.
How does this technology work?
The semiconductors used for energy conversion tasks are not usually made of silicon, but the option is generally gallium nitride (GaN) as it handles current better and reaches quite high frequencies. However, to maintain compatibility with existing manufacturing approaches where silicon is the raw material, most GaN devices are built on a silicon wafer that simply provides physical support rather than contributing circuitry.
For the researchers, silicon provided an opportunity: we know how to control its structure on an extremely small scale (nanometers) that could be used to place coolant channels right against the surface of GaN circuits.
The construction process is quite elaborate; Initially, extremely thin grooves are cut through the GaN and into the underlying silicon. Then a silicon-only etching process is used to widen them into channels with the original gaps through the copper-sealed Gan layer, which improves heat conduction into the water. Below these channels there is a set of alternate passages that act as feeders and sinks; cold water enters through a feed, circulates through the channel where it collects heat and is then extracted through a sink.
The researchers arrange things so that the hottest parts of the GaN portion of the device are placed very close to the channels, allowing for more efficient heat extraction. Many of the tests they have done involve testing different element geometries such as channel width and spacing, but the amount of energy required to force water through the system has been tested at all times, which also imposed another restriction in terms of geometry.
The best design they have found so far is capable of handling heat fluids of up to 1,700 watts per square centimeter while limiting the maximum chip temperature to 60ºC.
Refrigerating the die with water, is it the future?
To turn all this research into a usable device, the researchers have used a thick double-sided adhesive layer and cut channels in it with a laser and then the chip was glued onto this adhesive. The idea is to pump water into the adhesive, from which it will flow to the chip, and all of this packaged on a single standard electronic board with all relevant power and data connections.
With these data, the first functional prototype of a 176W TDP device required a flow of water in the die of less than one milliliter per second, limiting the temperature to a maximum of 60ºC, as we said before, showing that they have managed to overcome the barriers of the efficiency.
According to the researchers, in a data center around 30% of the energy consumed is for cooling systems; If this technology could be adopted on a large scale, it could greatly reduce the energy required for cooling systems, and not only that, but it would also significantly increase the speed and performance of current processors, since they are currently limited by temperature .
