Will Supercapacitors Come to AI’s Rescue?

In the United Kingdom, electricity provider National Grid faces a problem every time there is a soccer match on (or any other widely viewed televised event, for that matter): During halftime, or a commercial break, an inordinate number of viewers go to turn on their tea kettles. This highly British coordinated activity strains the energy grid, causing spikes in demand of sometimes thousands of megawatts.

In AI training, a similar phenomenon can occur every second. Because training is orchestrated simultaneously among many thousands of GPUs in massive data centers, and with each new generation of GPU consuming an ever-increasing amount of power, each step of the computation corresponds to a massive energy spike. Now, at least three companies are coming out with a solution to smooth out the load seen by the grid-to add banks of huge capacitors, known as supercapacitors, to those data centers.

When you have all of those GPU clusters, and they're all linked together in the same workload, they'll turn on and turn off at the same time. That's a fundamental shift," says Joshua Buzzell, vice president and data-center chief architect at the power-equipment supplier Eaton.

These coordinated spikes can strain the power grid, and the issue promises to get worse rather than better in the near future. The problems that we're trying to solve for are the language models that are probably 10 to 20 times, maybe 100 times larger" than the ones that exist today, Buzzell says.

AI workloads sometimes use power in short bursts, causing the power grid to experience a wildly oscillating load [blue]. Banks of supercapacitors can store energy on short time scales [purple], which compensates for AI's power bursts and creates a smoother power demand on the grid [red]. Siemens Energy

One solution is to rely on backup power supplies and batteries for charging and discharging, which can provide extra power quickly. However, much the way a phone battery degrades after multiple recharge cycles, lithium-ion batteries degrade quickly when charging and discharging at this high rate.

Another solution is to use dummy calculations, which run while there are no spikes, to smooth out demand. This makes the grid carry a more consistent load, but it also wastes energy doing unnecessary work.

Supercapacitors for AI Power Management

Several companies are coming out with a new solution using banks of supercapacitors. Supercapacitors have two parallel plates, like regular capacitors. But they also have an electrolyte layer in between the plates, akin to a battery. However, while batteries store energy chemically, supercapacitors store energy electrostatically, without the need for a reaction to occur. This allows them to charge and discharge quickly, offering a power backup on short time scales that doesn't significantly degrade over time. Siemens Energy is selling the E-statcom, a supercapacitor bank that is connected in parallel to the load at the level of the whole data center. They can charge and discharge at millisecond time scales, cycle up to 75 megawatts per unit, and promise to last between 12 and 20 years. The company is currently looking for customers in the data-center space.

Eaton is selling the XLHV (among other configurations), a supercapacitor bank the size of a single unit of a server rack (156 by 485 by 605 millimeters deep). One unit can dynamically deliver up to 420 kilowatts of power, and can also last up to 20 years. Some of these products are already deployed at data centers.

And Delta Electronics is selling the Power Capacitance Shelf, a 1-rack-unit-size bank of lithium-ion capacitors capable of supporting a 15-kilowatt load for up to 5 seconds. The capacitors Delta Electronics selected are kind of halfway between supercapacitors and lithium batteries," says Jason Lee, global product manager for supercapacitors at Eaton.

These and other such products can help smooth out load fluctuations on the grid. That's part of becoming a good grid citizen," Lee says. So instead of seeing all that fluctuation going back to the grid, we can take all the pulses and the low points and just smooth those out to where the utilities provide more or less average power."

This becomes particularly important when transitioning to renewables. The power supply of solar and wind power is more variable than that of oil and gas, as it depends on weather and other factors. This means it isn't always easy to ramp the power supply up or down in response to a varying load. Predictable loads allow for better planning and provisioning, which makes implementing some kind of smoothing mechanism particularly important.

However, Lee notes, this is not a panacea. The supercapacitors have a niche. They're not going to replace batteries everywhere. But these short-term events [like AI power bursts], that's an ideal application for supercapacitors."