Curbing AI power consumption in data centres

To learn more, Electronic Specifier’s Harry Fowle recently attended Nuvonix’s Onomen Michael Ojesebholo’s talk on the subject at the Engineering Design Show 2024 (EDS 2024).

Traditional data centres typically operate between 5-10kW, a number that is being dwarfed by the new wave of AI-focussed data centres cropping up today. These new data centres are often operating at three times the amount of traditional ones, reaching 30kW of power consumption as a minimum. This is a problem that Ojesebholo says, “calls for the next generation of power supply solutions.”

Environmental concerns

With such an increase in power consumption, one major problem comes to light – environmental impact. The impact that these new data centres are having cannot be ignored, with sources touting that these centres could account for 13% of global electricity consumption and 6% of the crabon footprint by the end of the decade.

The primary driving factors behind these statistics are the overall energy efficiency levels, as well as the cooling solutions that are built around them. As the power demands increases, so too do the footprint and thermal concerns. To address this, many data centres are employing water cooling solutions which, whilst effective, continue to drive up power consumption.

Wide bandgap is key

Wide bandgap semiconductor materials have become an integral part of curbing the power consumption of these new AI data centres, specifically silicon carbide (SiC) and gallium nitride (GaN).

These wide bandgap materials are something that Ojesebholo describes as being “revolutionary for power applications” in general – but especially for data centres. SiC for example has a “96% efficiency rate which can massively lower thermal losses,” he explains. “The power density gains are also a key driving force behind the adoption of wide bandgap materials.”

SiC technology in AI data centres has a wide array of benefits, such as:

• Zero turn-off tail current
• Normally off N channel MOSFET
• High-frequency operations
• High switching speeds
• Minimal switching losses
• Reduction of heat-sink requirements
• Maximum working temperature of 175C

GaN on the other hand is “just like SiC, but even better,” said Ojesebholo. This technology also has its own set of key benefits for AI data centres:

• Highly compact form factor
• Lower losses
• Improved thermal performance
• Even higher switching speeds
• Higher efficiencies than SiC

These materials are prime for use in the high-power density and thermal requiring environments which data centres operate. Ojesebholo continued: “These technologies can also reduce space, save on cooling elements and revolutionise form factors.”

Combining the two technologies

Whilst SiC on its own is great, and so too is GaN, the team at Nuvonix strongly believes that the ideal solution is a combination of the two technologies alongside supporting elements.

“GaN, for example, is something that has been spoken about a lot in the data centre space, but now we have witnessed a shift to a more combined approach,” explains Ojesebolo. GaN thrives on the AC side, whereas SiC more on the DC side, so why not use them where they’re strongest?

This approach can help to unlock the most efficient outcome, which after all is the main goal of these technologies being implemented in the first place.

Data centres require complete solutions

These AI data centres are calling for more than just efficiency gains, and to really thrive, a complete solution is needed. This approach includes factoring in various advanced components, smart systems, and power types.

With increasingly more complex solutions, systems are needing to become ‘smart’, even in the power side. “With smart solutions comes the need for MCUs, these devices certainly have a role to play now,” explained Ojesebholo. These MCUs have a range of key benefits in AI data centres such as:

• Dynamic load balancing
• Low power consumption requirements
• Predictive maintenance
• Security benefits
• Touch sensor capabilities

Then there is the need for more efficient power conversion methods, something that is being achieved with the use of GaN MOSFETs.

These power MOSFETS are driving three key areas within AI data centres:

• Low RDS(on)
• Low thermal impedance
• Fast switching speeds

Electrostatic discharge is also a bigger concern with these more advanced systems, for this reason, secure ESD protection plays an increased role. As Ojesebholo explained: “With more technical aspects and faster switching rates, ESD protection has never been more integral.”

The benefits of ESD in AI data centres include:

• Improved component safety
• Enhanced reliability
• High surge protection
• Fast turn-on
• Green part

Signal integrity amidst the noise of high switching and power density is also a bigger concern that must be addressed in these new approaches. Digital isolators can be used to ensure signal integrity and reliable communications across devices in a system.

The benefits of using digital isolators in AI data centres include:

• Improved signal integrity
• Enhanced safety
• High reliability
• High electromagnetic immunity
• Low power consumption

Finally, AC is being introduced increasingly into data centres in order to drive efficiency, which means there is a growing need for not just AC solutions but AC/DC solutions.

The benefits of this approach include:

• Compact and modular designs
• Overload and overvoltage protection
• High efficiencies of ~87%
• Built-in active PFC functions

Overall, the future of power-efficient AI data centres, as Ojesebholo puts it, “depends on robust, efficient, and intelligent power solutions.”