A Novel Method Utilizing Latent Heat to Effectively, Efficiently and Accurately Cool Microchips

Microchips are one of the most important technological developments in the 21st century, driven by the rise of AI. These devices need to be cooled to around 80 degrees Celsius to function properly, but as they get smaller and more powerful, cooling becomes a increasingly difficult task.

Currently, microchips are commonly cooled by cold plates, where liquid coolant flows through with an increase of temperature (see Figure 1). This can be expressed by:

(1)

: mass flow rate of water for this cooling method (kg/s)

: specific heat capacity (J/kg⋅°C)

: temperature increase of the water from inlet to outlet (°C)

Typically, water is used as the coolant because of its high specific heat capacity ( =4184 J/kg⋅°C).

After leaving the cold plate, the coolant then transfers heat to another loop of coolant through a heat exchanger. This second loop then transfers heat from all the microchips in the datacenter to a chiller, which rejects the heat outside of the datacenter (see Figure 1).

This cooling method has two main drawbacks:

As microchips become more powerful (producing more waste heat) and smaller (higher heat flux), this method becomes less effective.

This method requires a lot of water circulating in multiple loops.

A new cooling method is needed to provide effective thermal management for the next generation of microchips and datacenters.

Let us assume a flow of water absorbs the microchip's heat and changes phase from liquid to vapor. This can be expressed as:

(2)

: mass flow rate of water (kg/s)

= 2260x10³J/kg: latent heat of vaporization of water

Equating equations (1) and (2) gives us the following relation:

For =5°C, is equal to about 1%. Clearly, taking advantage of water’s high latent heat can significantly reduce amount of water needed. However, normally, water boils at 100°C, which is already above microchips' working temperature. To harness water's high latent heat while maintaining a low temperature, the pressure can be lowered to decrease the boiling point. The relationship between the two quantities can be shown through the August-Roche-Magnus formula (Figure 3):

(3)

: absolute pressure (ATM)

: boiling point (°C)

As shown in Figure 2, a vacuum pump can create a sub-ATM pressure , allowing water to boil at a low temperature inside the cold plate. I call this type of cold plate a micro-boiler. A hydrophobic net is placed in the micro-boiler, allowing steam to pass through and preventing liquid water from exiting. The steam is then taken away by the pump, which increases its pressure to ~1 ATM. The steam temperature increases approximately following the isentropic compression equation (Figure 4):

(4)

: steam temperature in micro-boiler (°C)

: steam temperature after passing through the pump (°C)

: pressure in micro-boiler (ATM)

: pressure after passing through the pump (ATM)

: ratio of specific heat of constant pressure to specific heat of constant volume

A high allows heat to be rejected efficiently from the loop through a radiator. The steam condenses in the radiator and returns to the micro-boiler through an expansion valve, reducing the flow pressure and temperature. The high temperature () could also allow for an energy conversion device to be added.

My project has successfully gone through two stages of testing. The 1st stage was the proof of concept. In the 2nd stage, the new cooling method was tested on a silicon chip in an open fluid loop, as shown in Figure 5. Additionally, the new method was compared to the existing cooling method. I am currently working on the 3rd stage: a closed loop prototype of the new cooling system. The results shown below are from the 2nd stage.

The open loop setup consists of:

Silicon chip: A 2cmX2cm silicon chip was attached to a heater that was created on a printed circuit board by milling its copper layer. The heater was powered with a DC power supply.

Micro-boiler: An acrylic chamber with an inlet for water and an outlet for vapor was constructed over the chip.

Vacuum pump: It together with a regulator was used to control the pressure inside the micro-boiler and remove steam to the ambient. (uncertainty: ±0.0005 ATM)

Data acquisition: A thermocouple was placed on top of the silicon chip to continuously record temperature ( ). (uncertainty ±0.5°C)

The new method was tested with chip powers of 15, 17.5 and 20W (uncertainty ±0.5%), which gave heat fluxes of 37500, 43750 and 50000 W/m², respectively.

Figure 6 shows the temperature recorded when the chip constantly dissipates 15W. After the chip thermally stabilized in the water filled micro-boiler at 1 ATM, the vacuum pump was turned on and regulated to lower the pressure by steps. Boiling was observed to become stronger inside the micro-boiler, which caused the temperature to decrease, quickly responding to pressure changes.

The temperature response and applied pressure were plotted in Figure 7 for all the tested powers. The trend is similar to that of the August-Roche-Magnus formula, with the chip temperature always being slightly higher than the saturation temperatures. Very clearly, a targeted chip temperature for a given chip power can be achieved by reducing the absolute pressure to a specific level.

How does the new method compare with the existing cooling method? To answer the question, comparative tests were conducted. As shown in Figure 5, a gear pump circulates liquid water through the micro-boiler with no phase changes. Because it is a closed loop, a radiator is used to reject heat. The same three chip powers and two water flow rates of 2475 and 5588 mg/s were tested. The chip temperature and water temperature were recorded.

The cooling performance of the two methods was compared using the heat transfer coefficient, which is calculated by:

: heat transfer coefficient (W/m²°C)

: heat flux (W/m²)

: temperature of the silicon chip (°C)

: saturation temperature for the new method or measured water temperature for the existing method (°C)

Figure 8 shows a significant difference between the two methods. In comparison to existing methods, the new method can achieve heat transfer coefficients 10 times higher, and with flow rates of water being lower by almost three orders of magnitude.

Water boiling had never been used as a cooling method for microchips. This project is making it happen (see Figure 9). A pressure-regulated cooling method utilizing latent heat of water was developed for the thermal management of microchips. In comparison with existing methods, the new method presents the following advantages:

The new method requires significantly less water.

The new method provides superior cooling performance.

The vacuum pump significantly raises the heat rejection temperature, which allows the heat to be recovered and harnessed as reusable energy.

The new method is able to provide accurately controlled cooling. The chip temperature shows a correlation with the regulated pressure. The correlation can be embedded in the feedback control shown in Figure 2.

Surpassing existing cooling methods in many aspects, the new cooling method will allow Moore’s law to continue and the chip node size to continuously decrease. The new cooling method will soon find widespread uses in high-performance computers and datacenters, reducing the overall size and complexity of the datacenter cooling infrastructure.

This project is continuing toward the final goal of commercialization:

I am optimizing and testing the method in a closed loop, as shown in Figure 10. The radiator and expansion valve need to be optimized.

The condenser will be connected to an energy conversion loop to recover and convert the energy from the cooling loop to reusable energy.

Once steps 1 and 2 are successful, I will contact microchip production companies to test the technology on their microchips.

Once step 3 is successful, I will contact major tech companies to implement the technology into their datacenters.

I would like to sincerely thank Dr. Jianxun Huang and Dr. Hangfei Duan at UBCO for their help in fabricating the micro-boiler and constructing the experimental setup. I would also like to thank Robert Roddie for organizing the Okanagan Shuswap Regional STEM Fair, which gave me the opportunity to come to the Canada Wide Science Fair.

Alduchov, O. A., & Eskridge, R. E. (1997). Improved Magnus` form approximation of saturation vapor pressure. https://doi.org/10.2172/548871

Alkrush, A. A., Salem, M. S., Abdelrehim, Osama., & Hegazi, A. A. (2024). Data Centers Cooling: A Critical Review of Techniques, Challenges, and Energy Saving Solutions. International Journal of Refrigeration, 160. https://doi.org/10.1016/j.ijrefrig.2024.02.007

Habibi Khalaj, A., & Halgamuge, S. K. (2017). A Review on efficient thermal management of air- and liquid-cooled data centers: From chip to the cooling system. Applied Energy, 205, 1165–1188. https://doi.org/10.1016/j.apenergy.2017.08.037

NASA. (2015). Isentropic Compression or Expansion. Nasa.gov. https://www.grc.nasa.gov/WWW/K-12/airplane/compexp.html

Peng, Y.-H., Wang, D.-H., Li, X.-Y., & Zhang, Y. (2022). Cooling chip on PCB by embedded active microchannel heat sink. International Journal of Heat and Mass Transfer, 196, 123251. https://doi.org/10.1016/j.ijheatmasstransfer.2022.123251

Sreejith Kochupurackal Rajan, Ramakrishnan, B., Alissa, H. A., Kim, W., Belady, C., & Bakir, M. S. (2022). Integrated Silicon Microfluidic Cooling of a High-Power Overclocked CPU for Efficient Thermal Management. IEEE Access, 10, 59259–59269. https://doi.org/10.1109/access.2022.3179387

Zhang, Q., Meng, Z., Hong, X., Zhan, Y., Liu, J., Dong, J., Bai, T., Niu, J., & Deen, M. J. (2021). A survey on data center cooling systems: Technology, power consumption modeling and control strategy optimization. Journal of Systems Architecture, 119, 102253. https://doi.org/10.1016/j.sysarc.2021.102253

Zhou, W., Dong, K., Sun, Q., Luo, W., Zhang, B., Guan, S., & Wang, G. (2022). Research progress of the liquid cold plate cooling technology for server electronic chips: A review. International Journal of Energy Research. https://doi.org/10.1002/er.7979