DPUs

DPUs, or data processing units, are purpose-built processors designed to offload data-plane tasks from host CPUs in data centers. They handle networking packet processing, cryptographic operations, and storage I/O offloads, enabling CPUs to run applications with higher throughput and lower latency. DPUs are typically deployed as PCIe cards or system-on-a-chip components embedded in NICs or programmable accelerators.

Architecturally, DPUs combine one or more general-purpose cores (often ARM or RISC-V), dedicated data-path engines for packet processing, DMA and memory interfaces, and hardware accelerators for cryptography and compression. They run software stacks such as DPDK and SPDK, and support programmable pipelines with languages like P4 and run-times compatible with eBPF.

DPUs differ from CPUs and GPUs in focus: they are optimized for offloading data-plane tasks rather than general-purpose computation or graphics, enabling better CPU utilization, security isolation, and lower latency for networked services.

Common use cases include virtualized network functions, firewall and VPN offload, TLS or IPsec offload, storage protocol acceleration (for NVMe over Fabrics), and policy enforcement in multi-tenant cloud environments.

DPUs gained prominence in the 2010s as data-center traffic and security requirements grew. Leading vendors include Nvidia with the BlueField family and Marvell with OCTEON DPUs, among others. The software ecosystem continues to evolve with DPDK, SPDK, eBPF, and P4 support, and some vendors market DPUs as successors to SmartNICs or as part of broader infrastructure processing platforms.