Abstract
A system is proposed herein for rhythm-predictive power management in artificial intelligence (AI) fabric network elements. The system learns the periodicity of AI training workloads from port counters local to a network device, predicts idle and busy windows, and schedules per-port power transitions so a port can enter a reduced-power state during a predicted idle interval and return to full capability before the next burst arrives. The mechanism requires no job, collective, scheduler, or external-controller knowledge. Confidence gating limits predictive action to ports with established periodicity and returns a port to reactive or always-on behavior when prediction confidence falls. By moving wake latency into the predicted idle window, deeper per-port power states can be used without imposing wake-up latency on burst-head traffic. Illustrative calculations show approximately 26% average per-port power savings under a representative one-second training rhythm, corresponding to about 65 kilowatts across 10,000 ports.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Calabria, Fernando; Alcaide, Juan; and Kosanovic, Nebojsa, "RHYTHM-PREDICTIVE POWER MANAGEMENT FOR AI FABRIC NETWORK ELEMENTS", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11777