High-Density, High-Efficiency Resonant Converters for Power Delivery to Modern Datacenters

Abstract

Global demand for computational power—driven by large-scale artificial intelligence, machine learning, and high-performance cloud workloads—is accelerating at an unprecedented pace. As GPUs continue to grow in power, size, and current demand, datacenter power-delivery architectures have evolved from traditional 12 V distribution to 48 V systems and are now transitioning toward high-voltage 800 V and ±400 V rack buses. Although these architectural shifts reduce copper mass, mitigate busbar losses, and improve system scalability, they also impose new requirements on the isolated and non-isolated DC–DC converters embedded throughout the power-delivery chain. Despite differing roles—from tightly regulated front-end converters to unregulated high-ratio intermediate bus converters (IBCs)—all stages share the same overarching mandate: maximize efficiency and power density under tightening electrical, thermal, and spatial constraints. To meet these demands, this dissertation investigates advanced methodologies for high-performance LLC resonant converters across multiple power-delivery stages. A unifying principle throughout the work is the deployment of PCB-integrated magnetics, which offer excellent manufacturability, repeatability, thermal performance, and ultra-low profile compared to conventional wound components. However, PCB transformers introduce stringent challenges in synchronous-rectifier (SR) termination, particularly when distributed output capacitors interact through parasitic inductances to create parallel resonances. These resonances can substantially increase transformer AC resistance at multi-hundred-kilohertz switching frequencies. Through analytical modeling, finite-element simulation, and hardware validation, this work characterizes these mechanisms and proposes optimized termination structures that suppress resonant peaking and lower conduction losses—recovering up to 3.5% full-load efficiency in 1 kW and 2 kW LLC prototypes. Building on this foundation, the dissertation next addresses the emerging high-voltage DC-distribution architectures for AI datacenters. A 6 kW, 800 V/50 V stacked LLC converter is developed using 650 V GaN devices in a three-level topology, enabling operation from both 0–800 V and ±400 V buses. A custom four-leg PCB-integrated magnetic structure merges two EI-core transformers into a compact, symmetric, and thermally optimized assembly. A top-cooled SR termination network is introduced to accommodate high output currents within severe footprint constraints. Operating near 600 kHz, the prototype achieves 98.7% peak efficiency and a record power density of 2070 W/in3—demonstrating the feasibility of compact GaN-based conversion for next-generation megawatt-class server racks. To support extreme-current GPU loads exceeding 1000 A, the dissertation also develops a high-density intermediate bus converter for emerging vertical power-delivery (VPD) architectures. A modular transformer unit-cell structure is introduced that maximizes magnetic utilization while enabling flexible series/parallel scalability. This concept is validated through an 840 W, 48 V/1.8 V LLC-DCX module achieving 2200 W/in3 and 95.5% peak efficiency, providing a viable pathway for shrinking regulator footprint, reducing PDN losses, and enabling back-side power delivery in future high-TDP GPUs. Finally, to reduce bulk-capacitor requirements in multi-kilowatt front-end AC–DC PSUs, the effective gain range of the LLC converter must be increased. To this end, a selective synchronous-rectifier (SR) phase-shift control technique is developed that extends the achievable gain without compromising soft-switching behavior. By phase-shifting only a subset of SR bridges in matrix-transformer structures, the required frequency variation is significantly reduced compared to prior-art methods, thereby minimizing circulating energy, lowering SR switching losses, and reducing the likelihood of ZVS loss in the primary switches. Experimental validation on a 3 kW, 400 V/50 V LLC converter—achieving 98.7% peak efficiency and 1300 W/in3 power density—confirms that the technique enables up to a 36% reduction in hold-up capacitance while maintaining stable and efficient operation. The approach is further extended to center-tapped rectifiers, where coordinated SR control eliminates reverse-conduction issues even when operating at the resonant frequency. Collectively, this dissertation advances the state of the art in high-density, high-efficiency LLC resonant conversion for modern datacenter power architectures. The contributions span device-level loss mechanisms, PCB-integrated magnetic design, high-voltage GaN converter topologies, extreme-current vertical power delivery, and advanced gain-extension control strategies. In addition, the work summarizes practical design guidelines for SR termination techniques for both center-tapped and full-bridge rectifiers across varying current levels, cooling configurations, and layout constraints. Together, these developments chart a clear path toward compact, thermally efficient, and scalable DC–DC converter platforms capable of meeting the escalating power demands of the AI era.