Power to GPU solutions is the final link connecting the power distribution network to the computing core in 800VDC data centers
In 800VDC data center architectures, Power to GPU solutions play a critical "last-mile" role. They are responsible for precisely and efficiently stepping down the intermediate voltage (such as 48V or 12V) from the high-voltage DC-to-low-voltage DC conversion stage to the ultra-low voltage (typically below 1V) required by the GPU core, while delivering currents up to thousands of amperes to the chip. They serve as the final link connecting the "power distribution network" to the "computing core."
Power to GPU solutions represent the "last hop" with an ultra-high step-down ratio, needing to convert the 48V or 12V bus voltage down to the GPU core operating voltage of 0.8V~1.2V, a step-down ratio as high as 60:1. The efficiency and precision of this stage directly determine whether the GPU can fully unleash its computing power. In addition, they must cope with extreme dynamic loads. During AI training tasks, GPU power consumption can fluctuate violently within millisecond timescales (e.g., from 100W to 1000W). Power to GPU solutions must possess ultra-fast transient response capability to ensure voltage droop is stabilized within ±5%, preventing the chip from downclocking or being damaged due to unstable power delivery. On the other hand, Power to GPU solutions must also handle extreme currents. Taking a single GPU with 1000W power consumption at 0.8V operating voltage as an example, its operating current reaches up to 1250A. This means the power stage must adopt a multi-phase parallel configuration (e.g., 16 phases or more) to share the current while ensuring current balance among all phases.
Power to GPU solutions also face many critical technical challenges, such as the adoption of multi-phase interleaved buck converters, with each phase carrying approximately 100A, reducing input and output ripple through interleaved control while providing redundancy, as well as the use of intelligent power stage modules that integrate the driver IC and power MOSFETs into a single package, shrinking PCB area, reducing parasitic inductance, and increasing switching speed.
Furthermore, Power to GPU solutions should be placed in close proximity to the load. To minimize voltage drops and losses caused by parasitic resistance and inductance on the PCB, the entire Power to GPU solution circuitry must be placed as close as possible to the GPU chip (typically on the backside of the GPU substrate or around its perimeter). The penetration of Gallium Nitride (GaN) technology is also an important trend. GaN devices, with their extremely low switching losses and higher operating frequencies, are progressively replacing traditional silicon MOSFETs and becoming the core choice for next-generation Power to GPU solutions.
Power to GPU solutions: delivering power accurately and efficiently to the GPU core
The final mile from the grid to the chip, and also the most critical mile, is to deliver power accurately and efficiently to the GPU core. The power consumption of AI chips continues to rise, placing unprecedented demands on transient response capability, current density, and voltage accuracy of power delivery. Power to GPU solutions is one of the core bottlenecks determining whether AI computing power can be effectively unleashed. They must provide stable, clean "last-mile" power to the GPU within millimeter-scale space, with extremely high efficiency, ultra-fast response speed, and massive current density. Their technological evolution is directly related to reducing the TCO (Total Cost of Ownership) of AI chips and increasing computing density. At the Arrow Electronics booth at PCIM Asia 2026, in the Power to GPU solutions segment, we will present the complete power delivery chain from high-voltage DC to the GPU core, showcasing complete solutions from multi-phase controllers, intelligent power stages to point-of-load converters, and how they drive next-generation AI computing with the highest density and efficiency.
In 800VDC data center architectures, Analog Devices' ADPM12160 (1.6kW) and ADPM12200 (2kW) IBC evaluation boards play the core role of "high-density intermediate bus converters (IBC)." They are key hardware implementations within the "Power to GPU solutions" stage, responsible for efficiently and stably stepping down the upstream 48V DC bus to the 12V voltage required by downstream GPUs, CPUs, and other loads, serving as a critical bridge in the entire power delivery chain. In a typical 800VDC power delivery chain, the front-end power stage (such as an SST) has already reduced the voltage to a 48V intermediate bus. These two IBC modules from Analog Devices enable efficient voltage conversion and power delivery for GPUs/CPUs, and feature high power density. They achieve performance breakthroughs in the 48V-to-12V conversion stage with 98% peak efficiency, kilowatt-level power density, and intelligent management functions.
Click here to view related information
In 800VDC data center architectures, the PGL6704.222HLT power inductor from Pulse Electronics, a Yageo Group company, plays an important "high-power-density energy storage and filtering" role in the final stage of "Power to GPU solutions." It is located in the 48V-to-12V step-down conversion stage, responsible for providing stable, clean front-end power to core chips such as GPUs and CPUs. As a key component of the buck converter, the PGL6704.222HLT inductor stores energy when the switching transistor is on and releases energy when the switching transistor is off, working with the output capacitor to smooth the pulsating voltage into a stable 12V DC. It supports step-down power up to 750W, making it an ideal choice for building high-power-density IBC modules capable of meeting the stringent space and efficiency requirements of AI servers.
Click here to view related information
In 800VDC data center architectures, the PGL7195.XXXHLT series TLVR inductor from Pulse Electronics, a Yageo Group company, plays a critical "transient response accelerator" role in the final stage of "Power to GPU solutions." It is located in the voltage regulator module (VRM) from 12V to the GPU core voltage (<1V), and is a key magnetic component specifically designed for the Trans-Inductor Voltage Regulator (TLVR) topology. With its dual-winding design optimized for TLVR topology, ultra-low inductance in the nH range, saturation current of hundreds of amperes, and extremely low DCR, it endows the GPU power delivery system with unprecedented transient response speed, ensuring that AI chips receive stable and efficient power supply even under extreme load fluctuations. It is one of the key technologies supporting the performance delivery of next-generation high-compute-power GPUs.
Click here to view related information
The Analog Devices ADP15193, specified as a 2.6V to 5.5V, 20A hot-swap E-fuse, is a low-voltage electronic fuse that plays the core role of "intelligent protection and power distribution hub in the 48V low-voltage domain" within the 800VDC architecture. Featuring high integration, high precision and programmability, fast response and recoverability, it is positioned at the final power distribution node after the high-voltage (800V) to low-voltage (48V) conversion and just before the final power delivery to loads like GPUs, responsible for providing precise and recoverable circuit protection for each load. Through precise fault protection, smooth hot-swap capability, and real-time system telemetry, it provides reliable protection for the "last mile" of power delivery from 48V to the GPU, and is a key component in achieving 7x24 uninterrupted stable operation of AI factories.
Click here to view related information
In 800VDC data center architectures, the onsemi NCP81295 hot-swap controller is located at the power distribution node after the high-voltage (800V) is stepped down to 48V or 12V via an intermediate bus converter (IBC) and just before final power delivery to loads such as GPUs, responsible for providing recoverable circuit protection and intelligent management for each load. It supports precise overcurrent and short-circuits protection, smooth "hot-swap" and inrush current management, and real-time telemetry and system management. It is an indispensable "intelligent power distribution protection node" in the low-voltage (12V/48V) power delivery chain of 800VDC data centers. Within the complete link from grid to GPU, onsemi positions it alongside AC-DC conversion, DC-DC power distribution, intermediate bus conversion, multi-phase control, intelligent power stages, and point-of-load regulation, collectively forming a comprehensive AI data center power delivery solution.
Click here to view related information
Article tags
