Meta’s paper describes a 150 MW data center hosting 83,000 GB200 GPUs and presents Dynamo Dimmer, which adjusts GPU power limits as a power-delivery device nears its ceiling.
This begins a distinct analytical record because no prior New Runtime record was supplied for this production-scale runtime power-control mechanism.
Thesis
Electrical capacity is becoming a runtime control variable for large-scale training, not only a facility-planning constraint.
Observed facts
Mechanism
When a power-delivery device nears its ceiling, the system reduces GPU limits collectively for the accelerators it governs, avoiding uneven local throttling that can create stragglers in synchronous training.
Why now
Meta’s paper presents electrical supply as a primary infrastructure constraint and describes runtime control as part of an end-to-end power-management process.
Implications
Cluster-wide coordinated power reduction may preserve training progress more effectively than isolated local throttling.
Performance trade-offs vary with workload and power limit, so operators need workload-specific measurements.
Evidence boundary
The reported measurements and operational claims concern Meta’s described 150 MW cluster and pretraining workloads; they do not establish equivalent outcomes for other hardware, facilities, or inference.
Watch conditions
Whether other operators report comparable coordinated power controls and workload-specific performance curves in production.