Power Budget Console

A power budget looks like a single number but is really two, and conflating them is the most common mistake in system power design. Peak power — every subsystem drawing its maximum at once — is what you must engineer the supply, the delivery network and the cooling to handle, because if the hardware can demand it, the infrastructure must provide it without sagging. Average power — the time-weighted draw under a real workload — is what shows up on the electricity bill. Size the supply to the average and it trips under load; budget the energy at peak and you wildly overspend.

The structure of an AI node makes one thing immediately clear: the accelerators dominate. Eight GPUs at several hundred watts each dwarf the host CPU, memory, networking and storage combined, so the budget is overwhelmingly a function of accelerator count and TDP. Everything else is secondary tuning. That's why the breakdown here puts the accelerator bar first and largest — it's where the budget is won or lost, and where any meaningful change must come from.

Sizing the supply adds two more factors. A headroom margin over computed peak — typically 15–25% — covers transient spikes, inrush and aging that a nameplate-exact supply would brown out on. And efficiency: even a top-tier supply loses a few percent of every watt as heat before it reaches the chips, so the draw from the wall exceeds the delivered load, and that loss is both energy cost and a cooling burden. This console folds both into the wall-draw figure.

Because every watt becomes heat, the power budget and the thermal design are one problem. The peak power here is the heat load your cooling must remove — verify it in the Junction Temperature and Power Density consoles — and the average power is the input to the energy bill in the Data Center Power console.