# Power supply and power shelf Canonical: https://semiaifoundry.com/learn/atlas/power-supply/ Revision: 2026-10-07 Rights: https://semiaifoundry.com/ai-atlas/licenses.html Power delivery · Atlas v1.3.0 # Power supply and power shelf Converts facility power into the regulated DC distribution used by the compute rack. [Explore in Atlas](https://semiaifoundry.com/ai-atlas/#power-supply)Rack shelf scale with multiple centimeter-scale supply modules Learning objectiveAccount for input power and conversion loss at a stated supply efficiency. ## The mechanism A power supply converts incoming electrical power into a form the equipment can distribute and use. Rack systems combine several supplies in shelves, sharing load and providing redundancy. A shelf’s installed nameplate capacity is not the same as the power continuously consumed by the rack, especially when redundant capacity is reserved. Efficiency also varies with load, and conversion loss becomes heat that must be removed. Facility input passes through supplies into DC distribution and then downstream regulation, with losses and capacity limits at each boundary. A failure scenario can remove one supply path while the remaining capacity carries the load, making reliability tangible. Supply counts and redundancy arrangements belong to the named rack configuration, not to a universal power-chain diagram. 1. Incoming facility power is converted into the DC supply used by the rack distribution system. 2. Multiple power modules share the load, with extra capacity arranged according to the required redundancy strategy. 3. DC power feeds downstream regulators, while telemetry and protection respond to changing demand or faults. ## Science and assumptions ### Conversion loss becomes a cooling load $$ \eta=\frac{P_{\mathrm{out}}}{P_{\mathrm{in}}},\qquad P_{\mathrm{loss}}=P_{\mathrm{out}}\left(\frac{1}{\eta}-1\right) $$ Efficiency is output power divided by input power. Conversion loss is output power times reciprocal efficiency minus one. Symbols and units η Conversion efficiency fraction P\_out Delivered output power W P\_in Input power W P\_loss Power dissipated inside the conversion boundary W Efficiency η relates useful DC output power to electrical input. The difference is dissipated in the conversion system. At large rack loads, even a small efficiency change can alter the heat that power shelves and facility cooling must remove. Installed redundant capacity is a separate concept: it changes the available power paths, not this conservation-of-energy accounting. Assumptions Steady average power at explicitly chosen input and output boundaries. Efficiency depends on voltage, load, and operating conditions; other rack and facility losses are outside a supply-only calculation. ## Engineering example Synthetic supply delivering 3 kW at 96% efficiency draws 3/0.96 = 3.125 kW and loses 0.125 kW = 125 W. That heat belongs to the supply boundary; it is additional to the 3 kW delivered downstream. ## Distinctions and tradeoffs Installed PSU capacity is not the rack’s constant power consumption. Redundancy reserves capacity, and actual load depends on equipment and operating workload. Redundancy improves availability but adds installed hardware beyond the active load requirement. Conversion efficiency, airflow, and service access affect the practical rack design. Large AI systems are constrained by deliverable electrical power as well as chip count. Power shelves turn facility capacity into a controlled supply that accelerators can actually consume. ## Specifications and interfaces DGX GB200 PSU example 5.5 kW per PSU NVIDIA DGX hardware guide configuration. Shelf example 6 PSUs / 33 kW per shelf Installed shelf rating, not a single processor’s power. DGX rack power Approximately 120 kW Published rack consumption example; configuration and operating workload matter. ### Interfaces Facility electrical input Input AC supply to conversion stages DC distribution Output Power to busbars and downstream regulators Telemetry/control Output Status, power readings, and fault information ## Check your understanding ### Why can a rack have more installed PSU capacity than its normal consumption? 1. Some capacity is reserved for redundancy and operating margin. 2. The unused rating proves the rack consumes no electricity. 3. Power supplies permanently store spare compute instructions. Answer and reasoning Some capacity is reserved for redundancy and operating margin. Redundant designs preserve enough capacity when a supply path fails or is serviced. The installed rating and active consumption therefore describe different things. ## Evidence and scope The efficiency identity is independent of brand. Values are synthetic, and real efficiency varies with input, load and temperature; distribution, regulator and facility losses must be accounted for at their own boundaries. No current calculator directly models this lesson’s main mechanism. Use the library for related bounds while retaining their stated omissions. ## Sources - [NVIDIA DGX GB rack systems hardware guide](https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html) - [NVIDIA DGX GB200 SuperPOD reference components](https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/dgx-superpod-components.html) [Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/power-supply-7f2ab40bff890e83.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html) ## Continue exploring ### Foundations - [Voltage regulator module](https://semiaifoundry.com/learn/atlas/voltage-regulator/) ### Related lessons - [Compute baseboard and tray](https://semiaifoundry.com/learn/atlas/compute-baseboard/) - [Baseboard management controller](https://semiaifoundry.com/learn/atlas/management-controller/) - [Rack power shelf and DC busbar](https://semiaifoundry.com/learn/atlas/power-shelf/)