Halley: A Predictive Control Plane for Optical AI Data Center Networks
AuthorsManav Mishra, Krishna M. Sivalingam
AffiliationsPrompeteer.ai, Los Angeles, CA 90066, USA · Department of Computer Science and Engineering, Indian Institute of Technology Madras, Chennai, IndiaEmails: manav@prompeteer.com, skrishnam@cse.iitm.ac.inCORRESPONDING AUTHOR: Manav Mishra
Halley anticipates failures in optical data-center networks and coordinates lasers, switches, and thermal controls to keep more AI training capacity online.
Key results
Composed simulation result, up from a 65.2% baseline.
Composed simulation under the paper’s lit-share interpretation.
At the assumed stress failure rate and correlation.
Estimated maximum under the exhausted-spare-window interpretation.
Fold reduction versus the spares-only plan under the stated cost model.
LUTs for the verified safety core.
What the paper found
Halley is a predictive control plane for optical AI data-center networks, including NVIDIA-style co-packaged optics and clusters at the scale Meta describes; the problem matters for large training runs such as Llama 3 and DeepSeek-V3. Its four controllers anticipate failures across the laser pool, optical circuit switch, microring modulators, and replaceable modules. LightReserve uses a beta-binomial model to size laser spares while accounting for correlated failures, which can impose a hard availability ceiling: under the paper’s stress assumptions, a 2× pool reaches only 99.670%. FoldFit packs traffic demands to schedule switch configurations, with a dwell-time guarantee within 11/9 of optimal under stated conditions, excluding switching delay. ThermaLead prepositions ring heaters using a one-step forecast, while TriageGuard assigns hot spares to the highest-risk modules; uncertain forecasts fall back to existing reactive controls. In composed simulations—not an end-to-end hardware trial—effective goodput rises from 65.2% to 76.5% when exhausted-spare windows count as lost, and reaches 67.4% when only dark-channel bandwidth is counted. The paper estimates up to 14.7% less data-center capacity for the same training under the first interpretation. Under its assumed cost model, reducing failure correlation before buying spares costs 5.3-fold less than spares alone. The safety-critical fallback and thermal-control circuitry is formally verified and uses 133 LUTs on the evaluated FPGA; the broader performance claims still depend on failure rates and correlations that require field measurement.
Original abstract
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. At a fixed spare ratio, correlated failure provably caps a pool's availability however large it grows. In this paper, we present a comprehensive predictive control plane framework, called Halley, that reads each resource's next state and acts before the failure occurs. It has four components, each carrying a formal guarantee: LightReserve that sizes the laser pool to provide a coverage bound, FoldFit that schedules the switch within $11/9$ of the optimal, ThermaLead that holds a microring with one-step-ahead control, and TriageGuard that protects the highest-risk modules first. Halley gives each layer a predictive controller that defers to the deployed policy when unsure. Composed from simulations at that point, effective goodput, the share of ideal throughput kept, rises from 65.2\% to 76.5\% if a window that exhausts its spares is lost. If only the dark channels' bandwidth is lost, it reaches 67.4\%. Per megawatt, that is up to 1.03--1.17 times the training; for the same training, up to 3.3--14.7\% less data center capacity. \textsf{Halley}'s fallback circuits are formally verified and are shown to achieve close timing on a commercial FPGA, and pass a 130~nm signoff.
Read the original paperMore in AI Hardware
Browse all 40 papers →PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?
Yue Zhang, Jilei Sun, Liqiang Jing, Yijie Wei, Ruohan Yang
PCBRouteAgent pairs an LLM’s board-level routing plans with KiCad’s routing engine to tackle complex PCB layouts.
QFabric: Guarded Cross-Domain Task Recovery and Its Limits on a Linux-RTOS Board
Ayushman Bhattacharya, Anwesha Chakraborty
QFabric tests whether tasks can escape timing trouble by moving between Linux and an RTOS, and finds that careful guards mostly prevent moves that would make performance worse.
A Hybrid GPU and Processing-Near-Memory Architecture with Shared DRAM for Efficient Prefill-Decode Disaggregation
Sanghyeok Han, Inseong Choi, Byungkuk Yoon, Gyeonghwan Park, Jae-Joon Kim
DuoPD pairs GPUs with memory-side processors to run LLM prefill and decoding together and improve inference throughput.