Data Processing Units (DPUs) are emerging as the third pillar of data center computing alongside CPUs and GPUs. NVIDIA's BlueField DPUs offload networking, storage, and security functions from the host CPU to a dedicated processor on the network card, freeing CPU cores for revenue-generating application workloads.
DPU Use Cases in Server Infrastructure
NVIDIA DOCA is the software framework for programming BlueField DPUs, providing abstractions for packet processing, storage virtualization, and cryptographic acceleration. In bare-metal cloud environments, DPUs handle OVS offloading, IPsec encryption, and NVMe-oF storage access without consuming a single host CPU cycle.
For multi-tenant hosting environments, DPUs enable infrastructure-level security functions that are invisible and inaccessible to tenants. Firewalling, intrusion detection, and traffic monitoring execute on the DPU, providing host-based security without installing agents in tenant virtual machines or containers.
Storage disaggregation with DPU-based NVMe-oF targets allows composable infrastructure where compute and storage scale independently. Remote NVMe drives accessed through the DPU appear as local devices to the operating system, delivering near-local latency while enabling flexible resource allocation across the data center.