Local edge GPU server — e.g. MIC-770 V3 with NVIDIA L4 (7,424 CUDA cores, 24 GB, 72 W) or MIC-75G30 dual-GPU module. Worthwhile at 10–20+ stations: correlate camera + thermal + vibration + energy locally, and serve as the second-opinion model for edge cascades.
Data out
The factory's local inference hub — stations talk to it, dashboards talk to it