The Additive Manufacturing of record-breaking pure copper heatsinks for high-performance computing applications

As the limits of computing power are continuously pushed in high-performance computing (HPC), effective cooling solutiofns have become increasingly crucial. The ability to keep processors cool during operation is one of the most critical factors for achieving peak performance and maximising equipment longevity. High-end graphics processing unit (GPU) clusters and Artificial Intelligence (AI) computing installations are often used in research environments and, at large scale, they are considered capital equipment, with maintenance and performance closely monitored.

Outside of large-scale installations, enthusiasts and professionals in the overclocking community constantly search for ways to extract every bit of performance from their hardware, pushing central processing units (CPUs) beyond their intended specifications. Overclocking increases clock speeds and power consumption, leading to significantly higher thermal loads that demand more advanced cooling solutions.

Additive Manufacturing and generative design have enabled the creation of the world’s most powerful metal additively manufactured liquid nitrogen (LN₂) heatsink for extreme CPU cooling. This groundbreaking innovation was developed in response to a request from the renowned overclocking experts at SkatterBencher and ElmorLabs.

Through this innovation, a new benchmark has been set in thermal management, and it demonstrates capability at the intersection of generative design software, advanced manufacturing, and materials science. This solution – or variants of it – can be applied to large-scale HPC applications using standard cooling fluids.

Background on the use cases

Though the average consumer may never require one, most of us are familiar with the fact that high-performance workstation PCs and scientific computing data centres are commonplace in engineering disciplines for various purposes including 3D modelling, numerical simulation, and 3D rendering, among many other applications. In these applications, power and thermal requirements significantly differ from the average consumer ultraportable, low-power email and web browsing machines.

Furthermore, in industrial data processing applications, and data centre and hosting applications, the need for large-scale rack-based computing systems creates a complex system of networking, storage, redundancy, power supply, and of course thermal solutions to manage the heat coming from hundreds or thousands of high-performance computing modules in one place.

The power density of data centres has been steadily increasing over the past decade, and this trend is expected to accelerate further. In 2017, the average power density per server rack was approximately 5 kW, but projections suggest that within just a few years this density will reach 100 kW or more per rack.

This increase in power density is a response to the growing demands of HPC and AI. It’s also forcing operators to reevaluate their cooling methods, as air cooling is no longer feasible in these power-dense environments. High-performance, specialised AI computing hardware is a fast-moving technical field with many of the usual chip makers entering the market to take part in a sector created by leaders such as Nvidia.

The high-performance CPU market is highly established, with transistor counts steadily increasing in line with Moore’s Law, while the thermal footprint of a CPU is now growing slowly after years of rising thermal design power (TDP). In contrast, the GPU computer market is technically advanced but, by design, allows for greater space, density, and overall hardware footprint, enabling a more complex subsystem. This architectural flexibility supports growth and innovation that significantly outpaces Moore’s Law.

New hardware domains – such as the tensor processing unit (TPU) and quantum processing unit (QPU) – are emerging and are expected to follow a similar technical development path to that of CPUs and GPUs, but at an accelerated pace. Furthermore, by 2025, annual spending on data centres to support AI applications is projected to reach $200 billion [1], marking a significant expansion over current global investment levels.

There is an associated issue with all this computing growth: where will all the power come from? With new, large-scale, power-hungry data centres rapidly emerging over the next two to three years, there will be a significant strain on the power generation capabilities of regional infrastructure. The challenge of modernising the power grid, embracing green energy, and addressing the climate impact of all this is inseparable from the growth of computing technology.

The record-breaking CPU cooling application featured in this article is an important example of what can be achieved by generative 3D modelling and Additive Manufacturing for high-performance computing applications. However, for the majority of the consumer or engineering workstation market, this concept, combined with LN₂ cooling, remains largely unobtainable and cost-prohibitive. Nonetheless, the demonstration is a strong indicator of thermal management capabilities that, with further development, could become viable for high-power density applications, such as those mentioned above.

From concept to design

Extreme CPU cooling using LN₂ is becoming the standard for computer overclocking enthusiasts as it allows for extremely low temperatures that are required to enable the desired extreme clock speeds. By leveraging the unique, two-phase generative design capabilities of Diabatix’s ColdStream platform, the engineering team has generated a heatsink that redefines the thermal performance of liquid nitrogen cooling. Two-phase cooling has particular advantages over single-phase cooling, which allow it to reach much higher heat transfer rates, making it an ideal choice for extreme CPU cooling.

Single-phase cooling

Single-phase cooling relies on heat transfer through air or a liquid coolant, such as water or oil, which remains in its liquid state throughout the entire process. Heat is transferred from a hot surface to the coolant by convection, which then transports the heat away from the heat source. This type of cooling is relatively simple to implement and, consequently, is widely used in many systems. Yet, it is limited by many factors, such as the coolant’s ability to absorb energy without heating up too much and the convective heat transfer coefficient values that can be reached at the solid-coolant interface.

Two-phase cooling

Two-phase cooling operates on a much more advanced principle by leveraging a phase change in the coolant. When a liquid coolant absorbs heat, it will change phase into a gas when it reaches the boiling temperature and heat continues to be added. The phase change allows the coolant to use the added heat as the latent energy required to evaporate the coolant. As a result, during the evaporating process, the coolant temperature remains constant and much more heat per unit of mass compared to single-phase cooling can be absorbed.

Due to the highly turbulent nature of the evaporation process, the convective heat transfer coefficients achieved at the solid-coolant interface can readily be an order of magnitude greater than those in single-phase cooling. This makes it highly attractive for high-power applications like data centres or power electronics. Once the liquid-gas mixture has carried the heat away, it is condensed back into a liquid, allowing the cycle to repeat. Because of the complexity of the phase change in both the evaporator and the condenser, this type of cooling is challenging to design using traditional design methodologies.

The Leidenfrost effect

In addition, a key factor in two-phase cooling that must be accounted for during the design process is the Leidenfrost Effect; a phenomenon where a liquid, upon coming into contact with a surface significantly hotter than its boiling point, creates a vapour layer that separates the liquid from the surface. This effect can completely nullify the effectiveness of two-phase heat transfer when it occurs, as the vapour layer acts as an insulating barrier and can effectively damage both the heat sink and the heat source. It is therefore crucial that two-phase heat sinks are designed in such a way as to prevent the Leidenfrost Effect during all operating modes.

The traditional design workflow for two-phase heat sink design is dominated by trial and error. A designer draws a cooling design, which can be tested against the specifications through either expert simulation software or prototype and practical testing. As with many other examples, when this design approach is applied to extreme CPU cooling, it results in a lengthy and costly process that demands multiple levels of expertise.

With the recent innovations by Diabatix in the field of two-phase coolant modelling, this trial-and-error process can be replaced by a highly effective generative design approach. This is an automated design process that requires minimal human input and interaction to achieve highly optimised designs that meet the functional requirements. By making use of physical modelling, massive computational resources, and state-of-the-art optimisation and artificial intelligence techniques, generative design can overcome the limitations of a traditional trial-and-error approach.

The simultaneous consideration of physics, design constraints, and manufacturing constraints is reduced to a few simple steps to formulate the input, with a generative design engine handling the rest. The starting point for generative design is not a best guess, but simply a description of the design target, an indication of the available design space, and a set of design limitations such as manufacturing constraints. When using generative design software, engineers are no longer designers in a committee, but they become managers of their own virtual design team.

The generative design input therefore reduces to the following five steps after which the process can start:

Key to this entire process is having a highly accurate physical model that describes the heat transfer from the CPU to the liquid nitrogen. By leveraging advanced two-phase Navier-Stokes equations in a conjugate heat transfer modelling context, ColdStream can deliver the required level of physical accuracy and efficiently handle phase change phenomena within the fluid flow. Furthermore, by incorporating both liquid and vapour phases in its calculations, ColdStream models the complex dynamics of boiling processes, including heat transfer and flow instabilities, while also automatically preventing undesired effects such as the Leidenfrost Effect.

Since no step in the process requires human intervention, the generative design approach eliminates human bias when transitioning from concept to manufacturing-ready design. While traditional design methods rely heavily on the engineer’s experience and intuition, which can inadvertently limit innovation, this limitation is removed with generative design. As a result, it has the potential to identify the optimal placement of material within the design space, which may not be immediately apparent or even counter-intuitive to human designers.

In practice, design engineers do not need to concern themselves with the complexity of physical modelling, as this is fully integrated into the process. Additionally, there are hardly any limits when it comes to model dimensions, operating conditions, or CPU powers. Instead, the simplicity of the process and the minimal input required ensure a high-quality user experience through an intuitive, easy-to-use interface (Fig. 4).

Production

This new cooling solution was manufactured utilising certified oxygen-free copper powder for superior thermal conductivity of 390 W/mK. This approach allows for geometries that traditional manufacturing methods cannot achieve to produce components optimised for performance (Fig. 5).

Maintaining the purity of the copper powder during the build is of utmost importance as any oxygen in the copper matrix has a detrimental effect on its thermal conductivity. 3D Systems’ vacuum chamber concept allows for a vacuum pre-cycle before the build job which actively removes air and moisture from the build chamber and the powder. After this cycle, the chamber is filled with high-purity argon gas.

This highly efficient and effective vacuum pre-cycle helps to achieve an extremely low oxygen environment. Furthermore, the vacuum chamber’s leak-tight design ensures that no oxygen can enter the build chamber, resulting in exceptionally low argon consumption during the build. This vacuum chamber concept helps to eliminate the risk of oxygen contamination in the powder feedstock, leading to stable powder chemistry and a significant improvement in the reusability of the certified oxygen-free copper powder batch.

Not only does this manufacturing process enable the above benefits, but it also provides the capability to manufacture parts with measured thermal conductivity which exceeds the measured values of parts made with other machines and the same powder.

Initial physical testing

Following its development, the LN₂ heatsink underwent initial physical testing by ElmorLabs to validate its performance. The detailed initial evaluation confirms the predicted heatsink’s exceptional cooling capabilities. In particular, the 3x faster cooldown behaviour compared to the Volcano LN₂ cooler, one of the top products in the market today, stood out during testing. Because the AI design and the Volcano have an identical mass of 1.7 kg and both use copper as the base material, this difference can only come from improved heat transfer at the surface and/or better heat transport inside the copper.

The tests demonstrated that the LN₂ container can easily handle a processor such as the Core i9-14900 KF with P-cores clocked up to 7.5 GHz, consuming 600 W of power. With a heatsink base temperature difference of 11°C compared to the Volcano LN₂ cooler, compared to a total heat up of only 9°C, the AI design proves to have the potential to carry more than twice as much power for the same base temperature. Therefore, the AI design passes the tests with flying colours with a huge margin left on reaching higher powers.

From the test data, the expected extraordinary thermal resistance of 0.011 kW was confirmed. That is only a 1.1°C temperature difference between the heatsink base and the liquid nitrogen per 100 W of CPU power, positioning it as one of the most powerful heatsinks in the world.
Further analysis revealed that this is only possible to achieve through an approximately +60% improvement of the heat transfer coefficient at the copper-LN₂ interface. It is clear that this leap in performance gain is only possible through, and is a direct result of, the power of thermal generative design.

Conclusion

The AI-designed LN₂ cooler demonstrated exceptional thermal performance, withstanding 600 W power at 7.5 GHz and achieving a thermal resistance of just 0.011 kW. Its optimised heat transfer coefficient at the copper-LN₂ interface allowed for a +60% efficiency gain over conventional designs. These results affirm that thermal generative design can surpass traditional cooling limits, positioning this heatsink among the world’s most powerful.

This pioneering project was initiated at the request of SkatterBencher and ElmorLabs, both leading authorities in the overclocking community. Their expertise and insights were instrumental in guiding the development process. Their ambition aligns very well with our commitment to pushing the boundaries of what is possible in thermal management. The current design is only the starting point of the collaboration. While preparing the supply chain for the commercialisation of the product, we are already preparing additional variants. At the top of the target list is to go beyond the 2,000 W mark in physical tests.

Authors

Scott Green
Principal Solutions Leader
3D Systems

Niels Holmstock
Team Lead, Industrial Application Development
3D Systems

Bram Hennebert
Application Development Engineer
3D Systems

Lieven Vervecken
CEO, Diabatix

Ine Vandebeek
Head of R&D, Diabatix

www.3dsystems.com
www.diabatix.com

References

[1] Bloomberg Professional Services, “Big tech 2025 capex may hit $200 billion as gen-AI demand booms”, October 4, 2024.