OPEN LINK ↗
// 59d agoRESEARCH PAPER
CMS updates AXOL1TL for real-time trigger anomaly detection
AXOL1TL is CMS's autoencoder-based anomaly detector in the Level-1 Global Trigger, scoring collisions on FPGA hardware inside a 50-nanosecond latency budget and keeping unusual events that may point to new physics. The 2025 CMS performance note documents the latest V4 and V5 deployments, including V5's April 2025 rollout and the newest real-data results.
// ANALYSIS
This is the kind of AI deployment that actually matters: the model only matters because the hardware and trigger stack can keep up with physics-grade latency.
- –The interesting part is the co-design loop, not the architecture alone: train on zero-bias data, compile to FPGA logic, then prove it still behaves in real time.
- –V5 looks like a real systems upgrade over V4, with self-supervised feature extraction and distributed-arithmetic optimization improving performance without blowing the resource budget.
- –The FPGA footprint is tiny by collider-trigger standards, at roughly 5% of GT LUTs and zero DSPs, which is exactly the kind of constraint AI infra teams should study.
- –For physics, model-agnostic anomaly triggering widens the search net without needing a signal hypothesis upfront.
// TAGS
axol1tlresearchinferenceedge-aimlops
DISCOVERED
59d ago
2026-03-29
PUBLISHED
59d ago
2026-03-29
RELEVANCE
8/ 10
AUTHOR
Better Stack