Enteligent launches modular power stack for AI racks

Enteligent has launched a modular three-stage rack-to-chip power conversion stack for AI data centres, designed for facilities moving to 800 VDC distribution.

Called PowerCore, the DC/DC conversion architecture carries power from an 800 VDC rack supply to the voltage rails used by GPUs. It consists of rack, server and GPU board components intended to give equipment designers a coordinated route for high-density power delivery.

The stack is built around three modules. RackCore is a 100 kW rack power shelf that converts 800 VDC to a 50/48 VDC bus. MaxCore is a server-mounted module that converts 50/48 VDC to a regulated 7 VDC rail. ChipCore is a 16-phase module placed near the GPU that converts the intermediate rail to 1.2/1.7 VDC.

The architecture is aimed at AI data centre operators facing rising rack power requirements. Enteligent cited industry projections that machine-learning systems will require more than 500 kW per IT rack before 2030, while NVIDIA is developing 800 VDC infrastructure for megawatt-scale racks.

That shift in power distribution is reshaping AI infrastructure design. As operators move to higher-voltage systems, the space taken up by low-voltage busbars, power shelves and transient buffering hardware is becoming a greater constraint inside racks.

Enteligent said its RackCore module removes the need for supercapacitor and battery backup shelves, reducing hardware requirements and heat at the GPU while freeing up rack space.

It added that the full PowerCore stack can deliver up to 50% greater rack compute density from facility power to the server. The company attributed that gain to the way the three stages divide power conversion into separate electrical domains that are optimised together.

System design

The design approach treats voltage, current, frequency, physical distance and electromagnetic emissions across the conversion chain as linked engineering factors. Enteligent said this system-level method is intended to reduce design complexity, rather than treating each stage independently.

The modules can be assessed as a complete chain or used separately with compatible downstream power designs. This gives server makers and AI infrastructure OEMs a choice between adopting the whole stack or integrating selected stages into existing designs.

The server-level MaxCore module uses what Enteligent described as a multi-phase, multi-path architecture. ChipCore, the final stage near the processor, provides the last point-of-load, or vertical power stage, for GPUs.

The launch comes as suppliers across the AI hardware market work to manage higher power densities and thermal loads in increasingly concentrated compute environments. Power delivery has become a more prominent part of rack design as operators seek to fit more accelerators into the same physical footprint.

Enteligent argues that at those power levels, the best architecture is not always the one with the fewest conversion stages. Instead, it said, the preferred design is the one that lowers the combined cost of conversion loss, conductor current, transient response, heat generation, electromagnetic emissions, filtering, board space and system certification from the facility bus to the chip.

Sean Burke, chief executive officer of Enteligent, said the company sees tighter integration between conversion stages as necessary as AI rack power climbs.

"As AI systems move toward much higher rack power, every conversion stage has to work as part of one system," Burke said.

"The PowerCore stack draws on our DC-to-DC experience in high-density power electronics and computing to give AI infrastructure and server OEMs a modular path from 800 volts at the rack to the GPU," he added.