AI Factory Operations and Control Foundations

Glossary

Building management system (BMS)

Facility software and controllers that run the mechanical plant, such as chillers, pumps and air handlers, commonly over protocols like ASHRAE's BACnet.

Checkpointing

Periodically saving a training job's full state to storage so that, after a failure, the job resumes from the last save rather than from the beginning.

Control plane

The set of systems, spanning facility and IT, that decide how compute, power and cooling are allocated over time in a data center. It is distinct from the data plane, which performs the computation itself.

Coolant distribution unit (CDU)

Equipment that pumps coolant through a liquid-cooled rack's cold plates and transfers the heat to the facility's water loop, forming the physical boundary between IT and facility cooling.

Data center infrastructure management (DCIM)

Software that tracks a facility's assets, space, power and cooling capacity and consolidates facility telemetry into one view of available headroom.

Demand response

An arrangement in which a large electricity user reduces its consumption when the utility or grid operator requests it, typically during periods of grid stress.

Goodput

The share of a cluster's computation that advances a result and is not lost to failures, restarts or redone work. It is stricter than allocation or utilization.

Workload orchestration

The software layer, such as Slurm or Kubernetes, that decides which jobs or services run on which hardware and when, and recovers them when hardware fails.