Designing power for data center racks
With higher voltage architectures moving large amounts of power over distance to lower current, different design choices can be taken to improve end-to-end efficiency. Unnecessary conversion stages tend to be removed where system requirements permit it, while each converter’s output range is typically optimized when matched to the operating window of the next.
However, the module with the highest peak efficiency quoted on its data sheet is not necessarily the best system choice. A fixed-ratio intermediate bus converter (IBC) can offer lower losses and higher power density, provided downstream regulators can tolerate its varying output. A regulated IBC maintains a tighter intermediate-bus voltage, although regulation can introduce additional losses. For the same input voltage and transferred power, a 4:1 IBC produces a higher intermediate voltage, reducing bus current and conduction losses. An 8:1 IBC produces a lower voltage, increasing intermediate-bus current but easing the step-down required at the final regulation stage. The best choice depends on many factors including the input range, load profile, board layout, cooling and the operating range of the downstream converters. Our website blog on choosing between a 4:1 and 8:1 conversion ratio provides further analysis.
Designing every component for an absolute peak can unnecessarily oversize the entire power chain. The challenge is to support peak and transient loads while avoiding excessive design margins, power instability and performance throttling. Power stages, control loops, decoupling and thermal paths, therefore, need to be assessed together. Power modules with clearly defined peak-power capability, parallel operation and digital control give power designers more flexibility to support transient loads.
Of course, cooling is part of power design too. Every watt lost in conversion becomes heat that must be removed from an already thermally constrained system, so it works best when cooling is designed together with power delivery from the outset. Top-side conduction paths, cold-plate compatibility and power delivery placed close to the processor can significantly reduce cross-heating and leave more PCB area available.
Embedded power solutions for AI data centers
Within 48 V architectures, power designers can choose between regulated and unregulated IBCs. Fixed-ratio products such as the BMR316 und die BMR321, with conversion ratios of 4:1 and 8:1 respectively, offer high conversion efficiency and power density when paired with downstream stages that accept a variable intermediate-bus voltage. By contrast, a regulated quarter-brick converter such as the BMR352 maintains a stable nominal 12 V rail as the input voltage varies. It also supports paralleling through active current sharing when more output power is needed. Peak-rated modules can handle brief load spikes above their continuous rating provided the duration and repetition rate remain within specified limits and the required cooling is maintained.
For next-generation AI power systems adopting 800 VDC or ±400 VDC distribution, the first need is an efficient conversion stage that can safely step high voltage down inside the rack, with isolation and protection defined at system level. We are developing DC/DC power modules for both architectures; for example, the BMR720. Initially designed for ±400 VDC systems, the BMR720 converts to a 50 V bus and provides up to 10 kW continuously or 15 kW at peak. This creates a route from emerging facility and rack standards into the lower-voltage domains required by server hardware without forcing each customer to create a high-voltage converter from scratch.
Near the processor, voltage regulator modules and vertical power delivery (VPD) become increasingly important. Solutions such as the BMR510 can deliver high current at output voltages down to 0.5 V, while custom VPD solutions place the final conversion stage directly beneath the processor. This shortens the high-current path, reducing resistive losses and improving transient response while freeing valuable PCB area. These benefits gain increasing value as processor current continues to rise.
As mentioned earlier, the product is only part of the power choice. Digital interfaces such as PMBus provide access to operating data, support configuration and help engineers understand faults or load behavior. Flex Power Designer, an industry-leading software tool for analyzing power-system performance under real-world operating conditions, lets teams model a complete digital power system before committing to hardware, comparing efficiency, thermal performance, phase spreading and droop behavior under expected conditions. Evaluation boards and application engineering then allow those assumptions to be tested before the design reaches the final platform.
Making every megawatt count
For every 1 MW handled by a conversion stage, each percentage point of loss represents around 10 kW of heat. Reducing that loss lowers the input power needed to deliver the same power to the processors and eases the cooling burden. It can also lower thermal stress within the power chain and leave more of a fixed site power allocation available for compute. Across multiple racks and conversion stages, those savings multiply.
In a power-constrained data center, conversion efficiency is essential. Additional generating capacity and greater grid capacity remain essential, but they take time and depend on factors outside the rack. Reducing losses from grid to chip offers a nearer-term way to deliver more of the site’s available power to the processors.
Power is the new currency of AI infrastructure, making efficient delivery and conversion a competitive advantage and commercial necessity. And companies that design embedded power and compute as one complete system will be better placed to turn every scarce megawatt into reliable AI capacity.
