Nvidia Unveils 'Scale-In' Security Layer for Agentic AI Workloads
Nvidia has introduced 'scale-in', a new layer in its AI factory stack that provides security across the entire east-west fabric, including memory, storage, and compute. This move is aimed at addressing the increasing complexity of agentic AI workloads.
The company's senior vice president of networking, Gilad Shainer, describes agentic AI as the most complex workload ever created, comprising a flow of operations that span multiple components, not just a single inference call. Scale-in is designed to handle this complexity by providing a secure and isolated infrastructure layer.
Nvidia's BlueField-4 DPU serves as the security plane in its AI factory architecture. It connects to every ConnectX NIC and manages 7.2 terabits of secure east-west traffic, outperforming the previous generation's 400-gig access pipe. The DPU runs out-of-band, collecting telemetry and enforcing policy without degrading GPU or storage performance.
The BlueField-4 DPU exposes its security engines, storage acceleration, and telemetry through the DOCA framework and DPU APIs to developers on a fully isolated device. This allows for greater control over agents running in the AI factory environment.