AI server AI power supply

AI server power supplies (PSUs) now range from 3 kW to 12 kW, leveraging Si, SiC, and GaN technologies to deliver high efficiency, compact form factors, and reliable power for compute-intensive AI workloads.

Power Ratings and Efficiency

Modern AI server PSUs are designed to meet the extreme power demands of AI accelerators and GPUs. Infineon's reference designs cover 3 kW to 12 kW, achieving >97% efficiency and power densities exceeding 100 W/in³ to fit within constrained server racks . These high-power PSUs support rapid load transients typical of AI workloads, ensuring stable operation during sudden spikes in GPU or accelerator demand .

Semiconductor Technologies

AI PSUs integrate silicon (Si), silicon carbide (SiC), and gallium nitride (GaN) devices to optimize switching performance, thermal management, and overall efficiency . SiC MOSFETs (CoolSiC™) and GaN transistors (CoolGaN™) enable higher switching frequencies, reduced losses, and smaller magnetic components, while traditional silicon MOSFETs (CoolMOS™) provide cost-effective high-voltage handling .

Architecture and Features

High-power AI PSUs typically use a two-stage architecture: a 3-level flying capacitor power factor correction (PFC) stage followed by a full-bridge LLC converter, providing compact, high-efficiency AC-to-DC conversion . Auxiliary energy buffer circuits offer 20 ms hold-up time and grid shaping during transients, improving reliability and compliance with data center power standards . Digital control is increasingly important. Microcontrollers with DSP capabilities manage real-time power distribution, predictive maintenance, and adaptive load response, ensuring efficient operation under fluctuating AI workloads . Solid-state relays replace mechanical relays for enhanced reliability and high-temperature performance.

Trends and Applications

AI server PSUs are evolving to support multi-phase designs and higher per-rack power, with projections exceeding 100 kW per rack as GPUs like NVIDIA's Blackwell family approach 1–2 kW per chip . These PSUs are critical for AI data centers, high-performance server racks, and cloud infrastructure, enabling machine learning, natural language processing, and other compute-intensive tasks .

Summary

AI server PSUs combine high power, high efficiency, and compact design using advanced semiconductor technologies. They feature robust digital control, hold-up capabilities, and architectures optimized for AI workloads, making them essential for modern data centers handling large-scale AI computations .