2021 Overview of China’s Data Processing Unit (DPU) Industry

The content originates from the report "China Data Processor Industry Overview (2021)" and focuses on the definition, classification, and technical routes of Data Processing Units (DPUs), as well as their industry chain, business models, main application scenarios, and global competition.

Early DPU products were supplied mainly by mature network equipment vendors and chip giants such as Mellanox, Netronome, Broadcom, and Cavium; later, smaller firms like BittWare and Ethernity entered the market as data traffic surged and CPU bottlenecks became evident.

Internationally, Nvidia, Intel, and Broadcom have launched multiple DPU products with transmission speeds up to 40 Gbps and storage I/O up to 32 Gbps, while Chinese domestic DPUs lag behind at 10 Gbps/8 Gbps, showing a 1‑2 generation gap in performance and reliability.

DPUs offload network‑stack algorithms and protocol processing from CPUs, freeing CPU cycles for other workloads; unlike traditional NICs, DPUs combine transmission and compute capabilities.

DPUs can be built on three core processor types—FPGA, multi‑core MP, and ASIC—resulting in product forms such as ASIC+GP (e.g., Nvidia) and ASIC+NP (e.g., Huawei), each offering different trade‑offs in performance, cost, and programmability.

Technology paths based on FPGA, MP, or ASIC enable vendors to meet diverse user requirements, balancing single‑point breakthroughs with broader market needs.

Rapid growth of data‑center traffic drives a shift from software‑based acceleration (SR‑IOV, DPDK) and embedded CPUs toward DPUs; FPGA and SoC are emerging as mainstream DPU technologies.

DPUs are positioned as a convergence of NIC and processor, providing both transmission and compute; Intel’s early software and embedded‑CPU solutions have been superseded by DPUs due to performance limits.

Supply‑chain considerations include upstream procurement of FPGA/MP/ASIC, with Chinese DPU vendors facing higher design costs and limited IP support compared to international counterparts.

The DPU market is still in its early stage, with high trial costs for smaller companies and a strong push from cloud giants like Amazon and Huawei, which acquire or develop DPUs for internal deployment.

Key application domains include data‑center networking (the largest market), telecom edge computing (driven by NFV and 5G), and emerging sectors such as autonomous driving, where each vehicle may require multiple DPUs to meet stringent bandwidth and latency demands.

Global DPU leaders—Nvidia, Intel, and Broadcom—maintain dominant market positions through acquisitions and technology integration; Chinese startups aim to differentiate through custom software services and price advantages.

Future trends point to 100 Gbps DPU development, continued migration of CPU‑intensive workloads to DPUs in edge clouds, and increasing importance of low‑latency RDMA and InfiniBand technologies, where Mellanox (now Nvidia) holds a leading edge.