CS Paper Radar

How AI Load Swings Reduce Datacenter Power Conversion Efficiency

Heetarth Joshi, Ali Alenezi, Fang Chen Lin, Prasad Enjeti, Xiaoyang Wang, Xin Chen

cs.AR eess.SY ·submitted 2026-10-23 21:11 UTC · v1

Abstract

AI training and inference workloads make GPU power swing repeatedly between compute and communication phases. As a result, the cascaded converters in a datacenter rack keep moving between operating points with different efficiencies. This paper uses an energy based, cycle averaged method to quantify that effect along a 240 V AC to 400 V DC to 51 V DC rack power path, using measured efficiency curves of commercially available GaN based PFC and LLC converters. For a 1 Hz, 50% duty-cycle load swinging between 5 kW and 510 W, conversion efficiency falls to 95.97%, compared with 96.82% at the equivalent constant average load, a loss of 0.88%. When 51/48V to 1V point of load conversion to the chip rail is included, it is shown that the loss grows to 1.66% (89.45% vs. 90.96%). These results show that efficiency rated at average load overstates real world performance, and that the penalty grows with each load dependent conversion stage between the grid and the GPU.

Closest papers

Nearest neighbours by meaning across the whole index.

AI Data Centers Meet Electrical Grids: A Review of Power Challenges and Coordinated Solutions

eess.SY ·2026-10-06 ·Mariam Mughees, Yuzhuo Li, Yize Chen, Yunwei Ryan Li
survey similarity 0.63
Abstract

Artificial intelligence (AI) workloads are creating large, power-electronic loads whose effects on power systems depend on location, temporal variability, controllability, and interactions with other loads and resources. Annual electricity consumption alone cannot capture these effects. This review examines how AI data centers affect power systems stability, renewable-energy integration, and low-carbon planning by tracing the power-delivery chain from accelerators and workloads through rack converters, uninterruptible power supplies (UPS), storage, and microgrids to the transmission point of interconnection. Reported oscillatory events and multi-gigawatt load transfers highlight how facility responses can amplify disturbances, while programmable workloads and controllable power converters offer opportunities for demand flexibility. Challenges and mitigation strategies are organized across device, rack, facility, and power systems levels, spanning millisecond-to-year timescales. A bandwidth-matching framework relates disturbances to the response capabilities of mitigation resources. The review distinguishes computational energy efficiency from power systems and sustainability outcomes: lower energy consumption per token does not necessarily reduce peak demand, electrical disturbances, or carbon emissions. It examines how coordinated workload scheduling, grid-interactive UPS systems, storage, high-voltage DC distribution, and flexible interconnection can support reliability, efficiency, and decarbonization. Key research needs include dynamic load models, high-bandwidth telemetry, carbon-and grid-aware scheduling, and mechanisms for valuing flexible computational demand.

0.08

Steady-State Equivalent Circuit Model for Data Center Loads

eess.SY ·2026-08-18 ·Muhammad Hamza Ali, Peng Sang, Hyeon Woo, Hyein Kang, Sungyun Choi +1 more
similarity 0.61
Abstract

Planners currently represent data centers as aggregate constant-PQ or ZIP loads in steady-state interconnection and contingency studies. These aggregate models are computationally convenient. However, they obscure the electrical relationship between computational workloads, server utilization, and grid-side demand. They ignore the internal power-electronic conversion stages of IT loads and assume homogeneous workload distributions across the compute clusters. This hides operating-point-dependent converter losses and efficiency variations. We propose a steady-state equivalent-circuit model (ECM) for data centers, which explicitly builds circuit models for IT loads, power supply units, cooling, and auxiliary systems. For power supply units, the equivalent circuit model explicitly represents internal power-electronic conversion stages. For IT loads, we develop a utilization-dependent server power model, and we combine it with loss-aware ECMs of power supply units. This approach captures the grid-side impact of heterogeneous workload distributions while preserving compatibility with conventional power-flow analysis. We evaluate this data center ECM in large-scale transmission power flows, using Monte Carlo simulations under heterogeneous and homogeneous cluster utilization. In comparison with the fixed-efficiency constant-PQ model, the ECM predicts that the most stressed line exceeds its thermal limit in about 30% of Monte Carlo samples. The results further show that homogeneous server utilization overstates line-loading variability by 17%-46% relative to heterogeneous server utilization, depending on the intra-cluster workload correlation.

0.47

Active Filter Design for Buffering Datacenter-Scale Power Fluctuations from Training AI

eess.SY ·2026-09-11 ·Dillon Jensen, Grant Wilkins, Obi Nnorom, Hugo Budd, Phil Levis +1 more
similarity 0.61
Abstract · 8 pages, 11 figures

Training large AI models requires thousands of processors working in parallel. This synchronized load poses a challenge for traditional power delivery architectures, because IT power ramps much faster than grid hardware can compensate. This paper presents a hardware solution to this problem, introducing a novel power delivery architecture which automatically detects changes in load power and compensates using on-rack energy storage, giving time for the grid to respond to the varying load. The architecture is validated experimentally, powering a training load while limiting the grid-side power ramp rate such that it stays within pre-specified range. The prototype presented is rated to deliver up to 10 kW of buffered power in a 400 VDC system, with a bill of materials cost of $3,500 USD. The system design is described in detail. A key advantage of this architecture versus other approaches is that it reliably buffers power fluctuations without requiring any changes to system software, making it compatible with any training workload.

0.22

Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States

cs.AI eess.SY ·2026-11-12 ·Alex C. Newkirk, Daniel L. Gerber, Erica R. H. Fuchs, Alex Hubbard, M. Granger Morgan +3 more
code/datahot direction similarity 0.60
Abstract · 80 pages incl. SI

The rapid growth of artificial intelligence is driving demand for new data centers at large scale and on short deployment timelines. To power them, firms connect to the commercial grid or contract generation directly, but interconnection and gas-turbine delays now incentivize on-site generation. We compare grid-connected and off-grid data center energy infrastructure (AC- and DC-coupled solar-plus-storage, islanded natural gas, and grid interconnection) across the continental U.S. Using an open-source technoeconomic framework built on empirically calibrated AI training loads, climate-driven cooling, stage-by-stage power conversion, and physics-based battery degradation, we model designs providing equivalent energy services. Our central contribution is a speed-adjusted cost of energy that values the compute unlocked by faster construction, making options with different deployment timelines directly comparable on one physically consistent basis. We find DC coupling reduces lifetime energy costs by 18% ($0.09/ kWh) on average versus AC counterparts by eliminating inverters and improving efficiency. Though rarely cheapest on unadjusted energy cost, DC-coupled microgrids' faster deployment makes them the lowest-cost modeled option in 85% of U.S. locations once foregone compute is valued, or 55% under accelerated natural gas timelines. For AI infrastructure, deployment speed rather than unit energy cost determines the economically preferred path to power.

0.31

From Grid to Chip: Power Architecture, Stability, and Flexibility of AI Data Centers

cs.ET cs.AReess.SY ·2026-09-10 ·Yubo Song, Rui Kong, Takuro Umihara, Pooya Davari, Frede Blaabjerg +1 more
hot direction similarity 0.58
Abstract · The manuscript has been submitted for review to IEEE Open Journal of Power Electronics

The rapid growth of artificial intelligence (AI) computing is transforming data centers into large, dynamic electrical loads. Their deployment is primarily constrained by energy availability and grid-connection capacity, which is further aggravated by the ability of power-delivery architectures, control systems, and computing workloads to operate reliably during fast grid disturbances. This article presents a technological perspective on AI data centers as grid-interactive computing systems. First, it reviews grid-integration bottlenecks, evolving connection policies, grid-code requirements, which has fostered new technological trends via spatio-temporal flexibility available through workload orchestration, cooling systems, on-site resources, and energy storage. Second, it maps the evolution of power-delivery architectures from medium-voltage grid interfaces to chip-level, discussing higher-voltage DC distribution, solid-state transformers, wide-bandgap devices, advanced chip-level power delivery, and liquid cooling. Third, it establishes a three-level stability framework spanning rack-level DC-bus dynamics, facility-level converter interactions, and system-level grid-coupled behavior. The framework connects dominant instability mechanisms, including constant power load effects, impedance interactions, forced oscillations, and operating-mode transitions, with suitable modeling, assessment, and mitigation approaches. Synthesizing these topics, this article highlights grid-to-chip co-design as a central requirement for scalable AI infrastructure, linking computing workloads, power-delivery systems, energy buffers, and grid operation.

0.27

Emerging Vertical Power Delivery in AI Data Centers: Challenges and Solutions

eess.SY ·2026-11-16 ·Charles Isaac Zahner, Amirhossein Habibi, Mehdi Ferdowsi
similarity 0.57
Abstract · 29 pages, 24 figures. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after …

The rapid growth of artificial intelligence (AI) workloads is increasing power demand in AI data centers, and processor current is projected to reach 2000 A and beyond. This trend is exposing the limitations of lateral power delivery (LPD) to processors and has motivated growing interest in vertical power delivery (VPD). This paper reviews the power delivery network (PDN) of AI data centers and the upcoming transition from LPD to VPD. First, the 800 VDC power delivery architecture, the impedance behavior of the PDN, and the limitations of LPD are discussed. Then, VPD implementation approaches, reported converter designs, and packaging technologies are examined, together with integrated voltage regulators (IVRs) and embedded passive components. In addition, power integrity simulations are presented to compare regulator and plane positions, from which a set of PDN design guidelines is derived. The results show that moving a voltage regulator module (VRM) from 50 mm away to directly below the die reduces the loop inductance by nearly half. In a complete PDN, this relocation lowers the low-frequency anti-resonant peak, whereas the mid- and high-frequency impedance depends on the decoupling stages and the position of the power and ground planes. This paper also shows that achieving practical VPD remains closely tied to challenges such as regulator height, passive-component embedding, and thermal management.

0.23

Environmental Assessment of GaN Current-Source and Voltage-Source AC-AC Variable Speed Drives

eess.SY ·2026-11-09 ·Luc Imperiali, Jonas Huber
similarity 0.57
Abstract · Submitted to an IEEE journal

Three-phase ac-ac converters for variable-speed drives can be realized conventionally as voltage-source converters (VSCs) or, leveraging emerging monolithic bidirectional GaN power transistors (M-BDSs), as current-source converters (CSCs). This paper presents a comparative life-cycle assessment of a realized VSC and a CSC demonstrators including grid-side and motor-side EMI filters, and with identical specifications (1.4 kW rated power, 200 V line-to-line rms grid and motor voltage). The embodied carbon footprints are similar (44.8 kgCO2eq for the VSC and 47.4 kgCO2eq for the CSC) and dominated by auxiliary and control electronics, and the printed circuit boards. Over 10 years of continuous operation under the NEMA Power Index mission profile for variable-load centrifugal systems (pumps, fans), the CSC's higher part-load efficiency translates into a clearly lower life-cycle carbon footprint of around 900 kgCO2eq compared to the VSC's 1130 kgCO2eq when operating with a carbon-intense electricity mix (440 gCO2eq/kWh). However, this difference almost vanishes when considering a future renewable electricity mix (25 gCO2eq/kWh).

0.24

Dynamic Modeling of Transmission Networks, Grid-Following Inverters, and UPS-Based AI Data Centers

eess.SY ·2026-11-02 ·Muhammad Hamza Ali, Amritanshu Pandey
similarity 0.56
Abstract

This document presents dynamic models of a transmission network and average models of power-electronic devices and loads in the synchronous $dq$ frame. We model the network in the positive sequence and its synchronous generators with the classical swing equation and a first-order turbine governor. The power-electronic models cover grid-following inverters and AI data centers. Each inverter has a phase-locked loop, a power controller, and a current controller. Each data center connects to the grid through two step-down transformers. Its low-voltage bus feeds an IT load, an induction-motor cooling load, and a constant-impedance load. A double-conversion uninterruptible power supply (UPS) serves the IT load. We model its rectifier, grid-forming output inverter, battery, and bidirectional DC-DC converter. A measured GPU power trace drives the IT demand. We replace every non-smooth operation with a smooth surrogate. This keeps the discretized model continuously differentiable for the nonlinear solver. We also show how we initialize each model at its power-flow operating point.

0.25