AI Server Power Efficiency Improvement: Key Role and Optimization Path of Magnetic Components

As artificial intelligence computing power demand explodes, AI servers have become a major energy consumer in data centers. The power consumption of a single AI server has risen from several kilowatts to tens of kilowatts, with rack-level power exceeding 100kW. Against this backdrop, even a slight improvement in power efficiency can yield significant energy savings and reduced operating costs. Magnetic components, which play a core role in energy conversion and transmission within power systems, have a critical impact on overall system efficiency. This article explores the key technologies and optimization paths for improving AI server power efficiency from the perspective of magnetic components.

The Importance of AI Server Power Efficiency

AI server power efficiency directly affects two core metrics of data centers: Power Usage Effectiveness (PUE) and Total Cost of Ownership (TCO). According to industry statistics, electricity costs account for approximately 30% to 40% of data center operating expenses, with this proportion being even more pronounced for AI servers due to their high power consumption characteristics.

From a technical perspective, AI server power systems typically consist of multiple cascaded energy conversion stages: from AC input to high-voltage DC bus, from the bus to intermediate bus voltage, from the bus voltage to on-board power, and ultimately to AI chip power supply. Each conversion stage incurs energy losses, and magnetic components, as core components in each conversion stage, account for a significant proportion of total power supply losses.

In high-efficiency power supply design, loss optimization of magnetic components is equally important as topology selection and switching device selection. With the widespread adoption of third-generation semiconductors (GaN, SiC), switching device losses have been significantly reduced, making the relative proportion of magnetic component losses increase, positioning them as a key breakthrough for further improving power efficiency.

Loss Mechanisms of Magnetic Components in Power Systems

To optimize the contribution of magnetic components to power efficiency, it is essential to understand their loss sources. The main losses of magnetic components can be divided into two categories: core loss and winding loss.

Core loss consists of hysteresis loss and eddy current loss generated by the magnetic core material under alternating magnetic fields. Hysteresis loss is positively correlated with operating frequency and flux swing, while eddy current loss is related to the square of frequency and the square of core thickness. In high-frequency application scenarios, core loss control is particularly critical. The operating frequency of AI server power supplies is evolving from tens of kilohertz to hundreds of kilohertz or even megahertz levels, imposing higher requirements on the loss characteristics of core materials.

Winding loss includes DC loss and AC loss. DC loss is determined by winding resistance and is proportional to the square of the RMS current. AC loss originates from skin effect and proximity effect, causing the equivalent resistance of windings at high frequencies to be significantly higher than DC resistance. In high-current output scenarios, winding loss often becomes the main source of loss in magnetic components.

Additionally, the losses of magnetic components are closely related to temperature. The loss characteristics of core materials vary with temperature, and winding resistance also increases with temperature. In the high-temperature operating environment of AI servers, the temperature rise effect of losses may further exacerbate, creating positive feedback that affects the long-term efficiency and reliability of the power system.

Optimization Strategies for Key Magnetic Components

In AI server power systems, different types of magnetic components perform different functions, and their efficiency optimization strategies also vary.

Optimization of PFC Inductors

Power Factor Correction (PFC) inductors are located at the power input stage, functioning to suppress harmonics and improve power factor. In high-efficiency PFC circuits, inductor losses directly affect light-load and full-load efficiency.

The optimization of PFC inductors primarily focuses on core material selection and winding structure design. Regarding core materials, magnetic powder core materials such as Kool Mµ and silicon iron offer distributed air gap characteristics, avoiding local overheating issues associated with concentrated air gaps while providing excellent DC bias capability. For windings, using flat wire instead of round wire can improve slot fill factor and reduce DC resistance; in applications with higher switching frequencies, Litz wire effectively suppresses additional losses caused by skin effect.

Optimization of Resonant Inductors

In LLC resonant converters, the resonant inductor together with the resonant capacitor determines the converter's gain characteristics and soft switching implementation. The loss of the resonant inductor directly affects the efficiency performance of the LLC converter.

The optimization of resonant inductors centers on core material selection and refined air gap structure design. Ferrite materials, with their high-frequency low-loss characteristics, are ideal choices for resonant inductors. However, fringing flux near the air gap can induce eddy current losses in windings, requiring optimization of air gap layout and relative winding position. Structural designs such as segmented air gaps and distributed air gaps can effectively reduce additional losses caused by fringing flux.

Optimization of Transformers

Transformers in power systems perform dual functions of voltage conversion and electrical isolation, with their losses significantly impacting overall system efficiency.

The core of transformer efficiency optimization lies in balanced design between core loss and winding loss. Core loss decreases as flux swing reduces, but reducing flux swing requires increasing winding turns, which in turn increases winding loss. Therefore, it is necessary to find the optimal flux swing operating point while ensuring the core does not saturate.

In terms of winding structure, planar transformer technology is increasingly being adopted in AI server power supplies. PCB windings enable precise control of turn-to-turn distance, reducing leakage inductance and winding AC resistance. Additionally, the good heat dissipation characteristics of planar structures help reduce temperature rise and maintain high-efficiency operation.

Optimization of Output Filter Inductors

Output filter inductors are located at the final stage of the power system, functioning to filter switching ripple and ensure load power quality. In high-current output scenarios of AI servers, the losses of filter inductors cannot be overlooked.

The optimization focus for output filter inductors lies in balancing DC bias characteristics and high-frequency impedance characteristics. Magnetic powder core materials have become the mainstream choice due to their excellent DC bias capability. For windings, copper strip windings or flat wire windings effectively reduce DC resistance and minimize conduction losses under high-current conditions.

Material Innovation and Process Upgrades

The realization of magnetic component efficiency optimization relies on continuous innovation in materials and processes.

Core materials are evolving toward lower loss and higher saturation flux density. Ferrite materials continue to reduce high-frequency losses through formulation optimization and sintering process improvements. Magnetic powder core materials such as Kool Mµ and high-flux materials are achieving increasing saturation flux density, providing more options for high-power-density applications. Amorphous and nanocrystalline materials demonstrate application potential in high-frequency, high-current scenarios due to their combined advantages of high saturation flux density and low loss.

Winding technologies including Litz wire, flat wire, copper strip, and PCB windings each offer distinct advantages. The development of automated winding processes has significantly improved the consistency and precision of winding parameters. The widespread adoption of processes such as vacuum impregnation and potting helps improve winding heat transfer performance, reducing the negative impact of temperature rise on efficiency.

Thermal management design is increasingly emphasizing synergistic optimization between magnetic components and cooling systems. Applications of thermally conductive potting materials and direct core-to-heatsink contact effectively reduce hot spot temperatures in magnetic components, maintaining low-loss states of cores and windings.

System-Level Collaborative Optimization

Efficiency optimization of magnetic components should not be confined to the components themselves but requires collaborative design at the system level.

The application of electromagnetic-thermal co-simulation technology enables engineers to accurately predict loss distribution and temperature fields of magnetic components during the design phase, thereby optimizing structural parameters of cores and windings. Through simulation analysis, loss hotspots can be identified and targeted optimizations performed to avoid efficiency degradation caused by local overheating.

Topology-magnetic component matching optimization is equally critical. Different power topologies have varying parameter requirements for magnetic components: LLC topologies are sensitive to resonant parameter accuracy, while phase-shifted full-bridge topologies have specific requirements for transformer leakage inductance. Through collaborative optimization of power topology and magnetic components, the performance potential of both can be fully realized to achieve optimal overall system efficiency.

Supporting optimization for wide-bandgap devices is a current industry hotspot. The high switching speeds of third-generation semiconductor devices such as GaN and SiC impose higher requirements on parasitic parameter control of magnetic components. Optimizing winding structures to reduce leakage inductance and distributed capacitance helps leverage the high-frequency advantages of wide-bandgap devices, further enhancing power density and efficiency.

Industry Outlook and Challenges

The improvement of AI server power efficiency is a continuously evolving process. Currently, the efficiency of mainstream AI server power supplies has generally reached above 96%, with some advanced solutions achieving peak efficiency of 98%. However, as power density requirements continue to increase, the difficulty of efficiency improvement grows accordingly.

The main challenges facing the magnetic component industry include: physical limits of loss control under high-frequency operating conditions, performance stability under high-temperature environments, and suppression of winding losses under high-current conditions. Addressing these challenges requires multidisciplinary innovation in materials science, electromagnetic design, manufacturing processes, and thermal management.

Looking ahead, the research and application of new magnetic materials, the development of integrated magnetic components, and the introduction of intelligent condition monitoring functions will provide new technological pathways for continuous optimization of magnetic component efficiency. In the wave of rapid development of AI computing infrastructure, magnetic components, as key enablers of power efficiency, will see their technical value increasingly highlighted.