AI workloads cause dramatic fluctuations in compute requirements, necessitating resilient, intelligent, and scalable power infrastructure for AI data centers. This includes implementing 3-phase uninterruptible power supplies (UPSs) that can handle substantial load increases and dynamic changes.
UPSs must be AI-tolerant. AI workloads are unpredictable, triggering power surges in milliseconds that place significant stress on supporting equipment. AI training algorithms are executed in sync on the graphics processing units (GPUs) in the data center, resulting in an oscillating load profile. The first line of defense against this stress is the UPS, so if the UPS isn’t designed to handle these frequent fluctuations, a load drop may occur.
Load tests confirm AI tolerance
Schneider Electric designs its UPS products to be able to handle various site and application conditions, which also includes the dramatic compute fluctuations of AI workloads. To that end, we’ve conducted rigorous load tests on our Galaxy V series UPSs, like the Galaxy VXL, designed specifically for demanding AI workloads.
The tests used load profiles of real hyperscale customers, so we didn’t rely on simulations. The tests showed that the Galaxy V Series UPS units handle fluctuations without load drops or performance issues. They use advanced double conversion and our patented eConversion topologies to handle the variability.
Tests were conducted at 40° Celsius ambient temperature and showed the UPSs comfortably supported loads of 1% to 100% with no limitations. Galaxy VXL, Galaxy VX, and Galaxy VL units also handled up to ten minutes of 125% overload, and one minute overload at 150% load level. These results confirmed the robustness of Galaxy V series for the AI loads.
Battery type matters for AI applications
The tests demonstrated that the type of battery the UPS uses makes a difference when supporting AI workloads. Frequent, dramatic fluctuations require fast discharge and charge rates for batteries.
Lithium-ion batteries outperform valve-regulated lead acid (VRLA) batteries when it comes to supporting the fluctuating demands of AI workloads. Especially above 110% load levels, VRLA batteries struggle with high-frequency load changes while lithium-ion batteries handle them seamlessly. This makes a critical difference for data center operators deploying AI at scale. Beyond performance, lithium-ion batteries offer a longer lifespan than VRLA options—reducing maintenance needs and lowering total cost of ownership when paired with AI-ready UPS systems.
Another important point to remember is that to support AI workloads, the entire power train needs to be AI-tolerant. Schneider Electric’s Galaxy V-Series alone can handle up to 110% load, eliminating the need for battery over-sizing. However, to prepare for loads exceeding 110%, the UPS should be integrated into a design and system that can handle dramatic fluctuations, including the switchgear, cabling, and circuit breakers. Schneider Electric offers end-to-end solutions for our customers’ needs in AI training and inference in terms of designing a robust power train.
Planning for redundancy and scale
When planning for future AI load scaling, data center operators must ensure their power protection systems provide resilience and scalability. UPSs such as the Galaxy VXL series offer parallel configuration options (n+0, n+1) to enhance redundancy and reliability, protecting AI workloads from power outages and load fluctuations.
While the future of AI remains unpredictable, resilience and scalability are non-negotiable for modern data centers. With solutions like Galaxy VXL, operators don’t have to build full capacity from day one. Instead, they can leverage the solution’s modular, scalable architecture, which allows operators to add power modules as their needs grow.
Switching modules is safer and easier with Schneider’s Live Swapcapability, which lets technicians remove and add modules without shutting down the UPS unit. Live Swap is like changing a tire without having to stop the car.
Discover more – AI-tolerant UPSs support for data center resilience
Ready to build resilience into your AI infrastructure? The Schneider load tests shows that the Galaxy VXL along with various other Galaxy V-series 3-phase UPS solutions maintain stable power output despite the high fluctuations of AI workloads. This means our solutions support AI clusters reducing risk of AI power disruptions, and thanks to their built-in modularity and redundancy, Galaxy VXL UPSs allow data centers to add resilience, scalability, and future-proof operations.
Add a comment