Executive Summary

• This article is adapted from a talk delivered by Danel Turk, Data Centre Solution Architect at ABB, at our recent AI summit, DCi Horizons, on building the AI-ready data centre. Danel walked us through the evolution of backup power and what UPS architecture needs to look at to survive the AI era.
• Rack densities have increased significantly over the last 15 years; on average, they have risen from 3 kW to 3MW, which complicates how sites protect and draw power.
• The industry is transitioning to direct current architecture to cut conversion steps to achieve a huge 5% boost in efficiency.
• Transitioning UPS protection from low voltage to medium voltage level allows operators to use 14 times less copper and protects the local utility grid from AI workloads.

Over the last decade or so, the industry has undergone a massive transformation; the rapid scaling has fundamentally shifted the first question that gets asked during site development.

Danel Turk explains:

“The data centres have really gone up from like 15 years ago, when we started a little bit bigger data centres. It was direct sizes, we talked like 3-kilowatt. Today, direct sizes, we almost talk in the three megawatt range… and that also means that the data centre sites have gone up from a few 100 kilowatts up to the few gigawatts… The power question pops up in every discussion as the first few questions, especially on the European side as well, but it’s the same all over the world. If there is availability for the power, basically you can start building your data centres.”

Direct Current Architecture

One of the most significant developments in UPS architecture is Direct Current architecture; it’s not a new concept, we’ve seen this crop up every few years, but now it’s becoming more popular. This is because modern materials and hyperscale densities mean Direct Current infrastructure is actually viable for widespread deployment.

The standard practice for several decades has been alternating current feeds running down to individual server racks. Driving extreme power through multiple AC-to-DC and DC-to-AC burns crucial energy that needs to be redirected elsewere. After all, energy efficiency is the current aim with grid constraints. So, moving to Direct Current Infrastructure translates into a 5% efficinecy gaina cross the entire data centre.

The roadmap to integration

This shift to integrate DC into infrastructure has three distinct deployment stages.

• Sidecar approach: This is in partnership with the Open Compute Project and hardware leaders like NVIDIA, this integrates localed DC network alongside the rack. For dense racks scaling up to 500 kilowatts this is highly effective and it can also be retrofitted into existing brownfield AC data centres.
• Rectifier unit generation: this approach utilises a central transformer and rectifier unit to feed a DC network downstrea; it relies on a proven track record: “The similar concepts and architectures are used also in the marine last 20, 30 years, all the bigger cruise ships where people are going and driving around are also based on these technologies. So it is industrialized, proven, and the technology is there to go ahead.”
• Solid-State end goal: Direct Current architecture efficinecy relies on solid-state transformers and breakers, whch is custom-built to handle huge compute workloads. This approach is highly efficient for racks of 1MW and above.

Protecting the grid from AI workloads

Dense AI workloads are highly erratic compared to the standard cloud applications and this can cause pressure on the grid. Right now there’s a major disconnect between what a grid epects and to how a machine learning cluster actually behaves.

Operators can dream of a perfectly predictable solid load curve but that’s not the case with AI training and inference; they create volatile power spikes. SO when a data centre site scales into hundreds of megawatts – or even gigawatts – transmission lines can’t handle the rapid surgers and drops without risking wider instability on the grid. To close this disconnect, modern backup power needs to be a shock absorber, so operators can use fast-response battery storage to “shave the peaks” and present a clean, stable load to the utility grid.

Shifting to medium voltage UPS architecture

Shifting away from low-voltage systems to centralised medium-voltage UPS systems will offer huge advantages to data centres scaling to 5 megawatts.

Medium voltage systems can be scalable up to 25 MW in parallele, has 98% efficiency, optimised for 50-100% of data centre workloads, needs 14 times less copper in cabling and instead of a stanard cycle, it has a 15-year lifetime cycle design.

Furthermore, these systems, instead of using hundreds of small low-voltage architeture enables operators to future-proof the facility. Danel explains, “That means also less labor and also the less maintenance as well… Optimising that and putting it to the one place instead of close to the white space, it’s a lot easier to make service and the maintenance. So benefit is clear there.”

The time of achieving efficient, future-proof UPS architecture is now; power electronics are evolving so rapidly and the sector itself is borrowing industralised, field-proven concepts from industries such as marine transport and railway sysrems (and being sucessful in doing so). So the tools are ready, all that’s needed is the foresight to design for them from day 0.