Halley predictive control plane for optical AI networks
Original title: Halley: A Predictive Control Plane for Optical AI Data Center Networks
Halley boosts goodput in AI data center optical networks.
The paper presents Halley, a predictive control plane for optical networks connecting thousands of accelerators in AI training data centers. Four components provide formal guarantees for laser pools, switch scheduling, microring control, and risk triage. Simulations indicate goodput increases from 65.2% to 76.5% with reduced capacity needs.
- goodput rises from 65.2% to 76.5%
- up to 14.7% less data center capacity
- formally verified on FPGA
From the original article
1 min read at huggingface.co
A frontier AI model is trained in an AI data center, on tens of thousands of accelerators joined by a dedicated network, and every training step ends with all of them exchanging data at once. Between racks, that network is optical, and each of its four layers, from lasers to switches, runs a fixed or after-the-fact policy. Lasers sharing an enclosure, cooling, and power can fail together.
By Manav Mishra, Krishna M. Sivalingam · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.