AI workloads are redefining power-delivery requirements inside modern data centres. Ultra-high current demands, increasingly vertical power-delivery pathways, and advanced thermal architectures mean today’s power engineer must understand far more than traditional DC/DC conversion.
This guide explains how 20 essential terms impact power systems, organised into three sections:
- Power-delivery architectures and topologies
- Control, protection and digital optimisation
- AI, cooling and system-level trends that influence power design
1. Core Power-Delivery Architectures Shaping AI Systems
Modern AI hardware consumes extraordinary power levels — often several kilowatts per processor — across complex, multi-stage conversion paths. Understanding the architecture behind this flow is the foundation of power design for AI servers.
HVDC – High-Voltage DC Distribution
DC bus voltages above SELV levels (typically >60 Vdc) used inside equipment to feed high-voltage DC/DC converters, improving conversion efficiency and supporting higher-current loads such as AI accelerators. Examples include ±400V and +800V. As rack power moves beyond 100 kW, distributing power at ±400 V or +800 V HVDC becomes an efficient choice. Lower distribution current reduces copper losses, cable sizes, and conversion stages before power reaches the server.
IBA – Intermediate Bus Architecture
Data‑center power scheme using a 48V or 12V intermediate bus feeding Voltage Regulator Modules. From the HVDC feed, systems typically transition into an IBA — a stepped approach where power is first converted to a stable intermediate voltage before being locally regulated. In AI servers, this intermediate stage is often 48–54 V, selected for both safety and efficiency.
DCX – DC Transformer
Isolated, fixed-ratio DC/DC stage providing efficient bus conversion at high power. A key enabler inside HVDC-based architectures, a DCX transfers power between voltage levels using isolation and fixed-ratio conversion. DCXs allow high-power, high-efficiency distribution deeper into the rack or server chassis before final regulation.
LLC – Inductor-Inductor-Capacitor Resonant Converter
High-efficiency resonant converter used in power supplies for low noise and high density. LLC converters are widely used at front-end or intermediate stages to achieve high efficiency over varying load conditions. Their soft-switching characteristics make them ideal for the demanding thermal profile of AI environments.
Voltage Regulator Modules (VRM)
Module providing precise regulated power to processors or ICs. AI accelerators require sub-volt power at hundreds or even thousands of amps. VRMs are the final regulation stage delivering this power directly to the xPU package (CPU/GPU/NPU/etc. – see section 3). Their transient response capability is one of the most critical performance factors in AI boards.
TLVR – Trans-Inductor Voltage Regulator
Advanced voltage regulation topology using coupled inductors for high-current CPU power. TLVR is a next-generation VRM architecture offering faster transient response and improved efficiency at high currents. As AI accelerators impose extreme load steps, TLVR designs are becoming increasingly essential.
VPD – Vertical Power Delivery
Power architecture delivering current directly from board edge to high-current ASICs or GPUs. To overcome the limitations of lateral PCB routing, VPD routes power vertically through interposers or package layers. By shortening power paths, VPD improves distribution efficiency and reduces IR drop — essential for high-current AI processors.
TDP – Thermal Design Power
Maximum sustained power a device dissipates under typical workloads. Power engineers must understand TDP because it defines the sustained thermal limit of each AI processor, shaping power budgets, module placement, and regulator density. Higher TDP means tighter coupling between electrical and cooling design.
CESS – Capacitive Energy Storage System
Local energy-buffering system using high-capacitance storage (e.g., ultracapacitors) to absorb or supply rapid load transients, stabilizing voltage during sudden current changes in high-performance power systems such as AI accelerator boards. By absorbing and releasing charge close to the load, the CESS reduces stress on upstream converters and stabilises the PDN.
PDN – Power Delivery Network
Hierarchical power delivery system. The PDN encompasses the entire electrical path — from rack feed through VRMs to the silicon power bumps. Designing a low-impedance PDN is essential for maintaining voltage stability and preventing performance degradation in AI workloads.
Together, these concepts form the structural backbone of modern AI power delivery.
2. Control, Telemetry & Protection in AI Power Systems
As AI accelerators draw highly dynamic and sometimes unpredictable current profiles, modern power systems rely on intelligent control interfaces, monitoring capabilities and robust protection schemes to maintain safe and stable operation.
PMBus™ – Power Management Bus
Digital communication interface standard for power converters and monitors. PMBus provides real-time configuration and telemetry for DC/DC converters. It allows power designers to monitor voltages, currents, temperatures, fault states and performance metrics across thousands of nodes in an AI cluster.
AVS – Adaptive Voltage Scaling
AVS allows the xPU (CPU/GPU/NPU/etc. – see section 3) to request precise voltage adjustments based on workload or silicon behaviour. This reduces power consumption, improves performance-per-watt and stabilises fast load changes typical of AI inference and training.
DLC – Dynamic Load Compensation
Dynamic Load Compensation stabilises converter output during rapid load transients by adjusting control-loop behaviour and applying feed-forward techniques. DLC helps prevent voltage undershoot and overshoot when AI accelerators switch from idle to full load within microseconds, ensuring the PDN and VRM remain within tolerance.
OCP – Over Current Protection
Protects converters, busbars and downstream devices from excessive current events such as short circuits or fault conditions. In AI servers—with multiphase VRMs delivering hundreds of amps—fast, coordinated OCP response is essential to prevent cascading failures.
3. AI, Cooling & System-Level Trends Driving Power Requirements
To design power systems for AI workloads, engineers must understand the compute and cooling forces that dictate electrical design limits. These system-level trends influence everything from transient behaviour to total rack power.
LLM – Large Language Model
AI model trained on vast datasets for generative or analytical language tasks. LLMs (such as GPT-class models) demand massive compute resources and therefore massive power. Their bursty, parallel workloads shape the transient characteristics that VRMs, PDNs and local energy storage systems must handle.
xPU – CPU / GPU / TPU / NPU / IPU / FPGA
Generic term encompassing all types of compute accelerators – CPU (Central), GPU (Graphics), DPU (Data), TPU (Tensor), IPU (Intelligence), and others – used collaboratively in modern AI systems.
HBM – High Bandwidth Memory
3D‑stacked memory delivering very high bandwidth for AI/HPC accelerators. HBM dramatically increases thermal density around the xPU and requires tightly regulated low-voltage power rails. Its proximity to compute die influences VRM placement and power-stage thermal constraints.
D2C – Direct-to-Chip Cooling
D2C provides liquid cooling directly to cold plates on the processor package. This allows dramatically higher TDPs, influencing how much electrical power the VRMs and PDN must deliver and how tightly thermal and electrical design must be coupled.
CDU – Coolant Distribution Unit
The CDU regulates flow, pressure and temperature within the cooling loop. Its performance directly affects allowable electrical load, VRM temperatures and system efficiency.
PUE – Power Usage Effectiveness
The primary data-centre efficiency metric – total facility power divided by IT equipment power. Improvements in converter efficiency, VRM design, PDN optimisation and liquid cooling all contribute to better PUE at scale.
Conclusion
The AI revolution has created a new environment where power electronics, compute architecture, cooling technologies and system-level optimisation are inseparable. Understanding these 20 foundational terms helps engineers with the knowledge they need to design and scale reliable, high-efficiency power systems for today’s increasingly demanding AI workloads.
As architectures evolve — with higher TDPs, denser PDNs, advanced VRMs and VPDs, liquid cooling and HVDC distribution — staying fluent in the language of modern power design becomes essential.
To continue building your expertise, we’ve compiled one of the industry’s most comprehensive and continuously updated technical glossaries. Why not bookmark the full glossary of technical abbreviations at Flex Power Modules for future reference to deepen your understanding and stay ahead of emerging power-design trends: