Diamond for the next generation of “cool” Artificial Intelligence

The Cooling Challenge in AI Data Centers

In 2019, I had the chance to visit a data center of NVIDIA in Santa Clara. While the packed and stacked arrays of computing equipment were impressive, what left an even deeper impression on me was the constant, chilly breeze from the air conditioning systems that was constantly blowing in my face. These cooling systems are crucial to prevent overheating, which can lead to equipment failures and reduced lifespan of critical components. The data center energy costs are mainly contributed by two major intensive processes: 1. Computing power and server sources with up to 40% and 2. Cooling systems with 38-40% of power consumption to dissipate heat. (1) (Image 1: left) High computing power requirements of generative AI is posing higher demands on the dissipation of heat, approaching the limit of conventional air cooling and heat dissipation. To sustain high-performance AI computing, innovative cooling solutions are necessary. This is where diamond comes into play.

Diamond - The champion in thermal conductivity

Diamond possesses extraordinary thermal conductivity, making it an ideal candidate for next-generation AI chips. At room temperature (300 K), single-crystal diamond has a thermal conductivity of 24 W/cmK, ranking first followed by copper with 4 W/cmK and 1.5 W/cmK for silicon. (2) In addition, Diamond has a high breakdown electric field, making it a strong aspirant for miniaturizing power electronics. These properties enable AI chips to operate at higher frequencies, with greater power efficiency and faster switching speeds. (3)

Image 1: a) Shares of data center power consumption. Replotted from data in (1) b) Thermal conductivity. Replotted from data in (2).

Single Crystal Diamond Wafers On Lab Scale

The semiconductor industry has long relied on single-crystal silicon wafers available thanks to the  Czochralski Process, invented in 1916. A cylindrical silicon single crystal („seed crystal“) is immersed into molten silicon (> 1400°C) , on which it begins to crystallize. While the seed crystal is slowly pulled out of the melt, crystallization continues at the interface. At the end of the growth a single crystal ingot is obtained, from can be sliced into wafers. From invention of Czochralski process to the first large scale commercial production of 100 mm silicon wafer in 1975, almost 60 years passed to mature the quality and reliability of the process, emphasizing the time span it could potentially need for the Diamond Wafer technology to mature and scale up.

Traditionally, High-Pressure High-Temperature (HPHT) technique was used to grow diamond crystals. However, a more scalable alternative, Microwave Plasma (MP) Chemical vapor deposition (CVD), allows diamond to be grown using precursor methane and hydrogen under sub-atmospheric pressure. The challenge with CVD is that it often leads to polycrystallization at the edges, limiting wafer size. (4)

The homoepitaxial growth of diamond by CVD on predefined substrate such as HPHT diamond substrates leads to single crystal diamonds. However, substrates from HPHT are limited to smalls dimensions.

To overcome this, scientists have explored heteroepitaxial growth - growing single-crystal diamond on a non-diamond substrate. (5) Iridium surface have emerged as the preferred choice, with a technique called Bias Enhanced Nucleation (BEN) proving particularly effective. (6) Covered by a 1 nm thick carbon matrix, epitaxial diamond nuclei are formed. Single primary nucleation buried the epitaxial island can expand laterally, during which isolated secondary nuclei are continuously generated. This breakthrough Ion Bombardment Induced Buried Lateral Growth (IBI-BLG) has enabled lateral expansion of single-crystal diamond, achieving wafer diameter of 90 mm and a weight of 155 carats.

Image 2: Large-size diamond wafer by IBI-BLG process. Reproduced from (6). 

From Lab to Large-Scale Manufacturing

This scientific breakthrough was further refined by the German spin off Audiatec, which successfully developed wafer-scale single-crystal diamond using heteroepitaxy. (7) In 2021, Diamond Foundry Inc., a California-based company acquired this technology and publicly reported the first 100 mm mono-crystalline diamond wafer (110 carats). (8) Their approach involves a refined heteroepitaxy process to deposit carbon atoms controlled onto scalable substrates. The primary challenge now is to scale up the purity and crystallinity of these wafers to make them viable for established semiconductor manufacturing processes.

Application of Diamond for AI

Companies are already leveraging synthetic diamond technology to revolutionize high-power microchip cooling. One such example is Akash Systems, which has developed a method to attach polycrystalline diamond to GPUs using adhesives. This allows heat to be extracted directly from the GPU to a liquid cooling systems, improving thermal management. In a demonstration comprising an NVIDIA RTX 4070 GPU with and without diamond cooling, the GPU temperature was reduced by 50-60%, which would allow to increase the compute capacity by 200-400%. (9)

Conclusion: The Future of Diamond in AI Hardware

As AI computing power scales up, conventional cooling methods reach their limits. Diamond’s unmatched thermal conductivity make it a promising material for next-generation AI chips and power electronics. With continued advancement in wafer-scale diamond production, the next revolution in semiconductor technology is going to take off soon.