How our simulator estimates a data center's power
Pods, activity and outdoor temperature in; compute, cooling and the power system out. What the model does, and where its numbers come from.
The bryo power simulator estimates how much power a data center draws, and what its power system must carry, from a few inputs you can change. This post explains what it does and where its numbers come from.
What you set
- Pods and rack type. One pod is 7 racks. The rack types are the power settings published for liquid-cooled Vera Rubin NVL72 racks (156 to 227 kW per rack), plus GB300 NVL72 and a few others.
- Activity. How hard the GPUs work, from idle to full load.
- Outdoor temperature. It decides how much of the heat dry coolers can reject, and how much the chillers must trim.
- Power equipment. Grid or off-grid, generators, batteries, and how many transformers each block has.
What it calculates
- Compute power from the racks and the activity.
- Cooling power: pumps, fans, dry coolers and chillers. Most of the heat (96 % in a published reference design) leaves through the liquid loop; the rest goes to the air.
- The power system. An AC power flow (with the open-source library pandapower) checks every transformer, switchgear lineup and voltage. In auto mode, each block gets the smallest number of units that carries the load with its spare units out.
- A grid outage. Batteries and generators take over; the model cuts load where any component would exceed its rating.
Where the numbers come from
Every number in the model is labelled sourced, derived, proxy or assumed, and each sourced number points to a public document. Some are not published anywhere, such as how a rack’s power changes at partial load. Those are labelled as proxies or assumptions.
The results are estimates from stated assumptions, not a design. They help you see early which parts of a site set its limits, before you talk to equipment makers.
Open the simulator and try your own site.