Halley: A Predictive Control Plane for Optical AI Data Center Networks

What happened: Halley is a predictive control plane for optical networks in AI training clusters that forecasts failures across four layers, from shared laser pools to optical switches, and defers to the deployed policy when predictions are uncertain. Key numbers: Effective goodput rose from 65.2% to 76.5% when spare Why it matters: The paper states that at a fixed spare ratio, correlated laser failure provably caps a pool's availability however large it grows, and Halley's components each carry a formal guarantee. Its fallback circuits are formally verified and pass a 130 nm signoff.

Halley is a predictive control plane for optical networks in AI training clusters that forecasts failures across four layers, from shared laser pools to optical switches, and defers to the deployed policy when predictions are uncertain.

The paper states that at a fixed spare ratio, correlated laser failure provably caps a pool's availability however large it grows, and Halley's components each carry a formal guarantee. Its fallback circuits are formally verified and pass a 130 nm signoff.