RBF hybrid edge-HPC architecture enables low-latency AI inference with delayed updates

Emerging cyber-physical systems—like smart agriculture, autonomous vehicles, and industrial IoT—require real-time AI inference from streaming sensor data, but the models behind them often depend on high-fidelity simulations run on remote HPC clusters with batch scheduling. This creates a fundamental latency mismatch: edge devices need instant decisions, while model updates are delayed by queue times and simulation throughput. A new paper from researchers at UC Santa Barbara and collaborators introduces RBF (Reverse Backfill), a hybrid architecture that reinterprets HPC backfilling—normally used to fill idle slots with lower-priority jobs—to instead prioritize opportunistic computation that improves model accuracy.

RBF deploys lightweight surrogate models on edge devices (including private 5G, cloud, and local hardware) for immediate inference, while asynchronously incorporating improved models from HPC as they become available. The architecture is pluggable, supporting different surrogate models and orchestrating across heterogeneous infrastructure. The team validated RBF in a real-world digital agriculture deployment, coupling edge sensors with computational fluid dynamics (CFD) simulations to infer airflow patterns in a large agricultural screenhouse. Their evaluation quantifies simulation latency, training cost, inference throughput, and the impact of delayed model updates on prediction accuracy. Results show RBF maintains continuous, low-latency inference while progressively improving model fidelity, even with irregular and delayed updates from batch-scheduled HPC systems.