SCADA

Overview

SCADA stands for Scaled Accelerated Data Access — a GPU-initiated storage interface presented at the 2025 OCP Global Summit. It scales to 100K threads instead of 100 and automatically routes data via GPU-to-GPU communication down to the drives, instead of relying on manual partitioning. Because memory fabric tends to break at scale, SCADA introduces an API that supports errors and retries, letting a disc be taken offline and brought back in, with everything GPU-initiated — GPUs cannot run a file system and there is no CPU concurrency available to ask for help. SCADA enables applications such as cuGraph WholeGraph (for GNNs) and cuVS (for vector search and indexing) to be left unchanged while gaining a GPU-side specialized cache that bundles requests to the storage server, exposing user, tiered, and backend-storage views.

(Note: "SCADA" here refers to NVIDIA's Scaled Accelerated Data Access project; it is unrelated to the industrial-control SCADA acronym.)