Born of the post-Moore era and heterogeneous computing, the DPU (data processing unit) has quickly set the chip industry alight. Nvidia recently announced that its DPU lifts IOPS between servers and storage above 41 million – more than four times the previous world record – raising the ceiling on data-centre storage access. Chinese DPU design firms are raising money fast: Zhongke Yushu closed a nine-figure A+ round in December 2021, and Xinqiyuan and Nebula-linked startups each raised hundreds of millions of yuan in 2021.
A new opportunity for advanced computing
As Moore’s law slows, the marginal cost of CPU performance gains climbs – CPU performance grows only about 3% a year, while global data volume compounds at nearly 50% annually and demand for compute doubles roughly every four months, by IDC figures. Data centres face enormous compute and bandwidth pressure, and that is where the DPU comes in.
Its most direct effect is to relieve the CPU: taking over infrastructure functions such as network, IO and storage virtualisation that consume compute without contributing directly to applications, freeing CPU cycles for the workloads that matter.
Evolved from smart NICs, the DPU emphasises network processing: it connects to the network directly, its IO bandwidth matching network bandwidth, so protocol processing can be offloaded entirely. Wang Peng, strategic development VP at Xinqiyuan, told China Electronics News that the DPU’s essence is carrying network-specific stack algorithms and protocol computation, its core value freeing CPU resources so other compute modules process business data efficiently. Network protocol processing, data security and algorithm acceleration – problems that ‘the CPU does poorly and the GPU cannot do’ – are its future territory. The CPU-GPU-DPU combination, he said, is the next leap in collaborative processing.
Zhang Yu, senior VP at Zhongke Yushu, called the DPU a milestone in the shift from general to specialised heterogeneous computing: using software-defined approaches to pool infrastructure resources and handle networking, storage, security and QoS more efficiently, the industry has positioned it as the data centre’s ‘third mainstream chip’ after the CPU and GPU.
Nvidia and Intel lead the contest
Nvidia, Intel, Xilinx, Broadcom and Marvell are all positioning in the DPU market, setting technical standards and building product ecosystems.
Nvidia moved first on both hardware and software, as it once did with GPUs. It launched the DPU and the DOCA software framework in 2020; BlueField-2 delivers 200 Gb/s and accelerates security, networking and storage tasks including root of trust, key management and RDMA; BlueField-3, released in 2021, pushes software-defined networking, storage and security to 400 Gb/s. DOCA abstracts and encapsulates BlueField’s hardware capabilities behind programmable interfaces, an open platform for software-defined, hardware-accelerated applications.
Intel’s IPU, released in June 2021, is its answer: taking over infrastructure functions such as storage and network virtualisation to free CPU cycles. The Oak Springs Canyon IPU carries two 100 GbE interfaces with virtual switching, NVMe and RDMA support plus hardware encryption; the Mount Evans IPU offers 200G bandwidth for cloud-scale processing. Behind it sits the oneAPI framework, giving developers a unified standards-based programming model and easing code porting into Intel’s heterogeneous ecosystem.
Both are building ecosystems: Nvidia says ASUS, Dell, Fujitsu, Gigabyte, Inspur and Lenovo plan to integrate its DPUs; Intel announced partnerships with Inspur, Ruijie Networks and Silicom in December 2021 to co-design new IPU solutions.
Huge domestic application space
Domestic funding is active alongside the international contest: Zhongke Yushu’s A+ round funds chip R&D, mass production and market development; Xinqiyuan closed Pre-A3 and Pre-A4 rounds for its next-generation DPU; Nebula ZhiLian raised an angel round to accelerate DPU development.
The application base is the ‘geographic advantage’. China’s traditional IDC market, the DPU’s main direction, is growing double-digit – IDC projects RMB 161.26 billion by 2023. DPU also prospects well in 5G and data security.
In communications, network virtualisation is a key 5G technology, and NFV sharpens operator demand for edge computing – strict bandwidth, latency and reliability requirements that stimulate DPU demand. Zhang Yu expects 5G edge cloud, autonomous driving and vehicle-road cooperation, and financial computing to grow faster than public cloud in DPU demand. In security, the DPU can consolidate scattered OS-level security policies into hardware, funnel network and application data through the smart NIC for unified protection, and provide path monitoring and traffic visibility – layered protection from data transmission to user data.
International vendors lead on product maturity, but the DPU remains an early blue-ocean market: domestic and foreign players alike are at the starting stage, domestic demand is enormous, and DPU design is application-driven. Vendors need to land technology early in target fields and polish products and services, building a complete application ecosystem. Zhongke Yushu is expanding into telecom, finance, energy and internet data centres plus cloud-native acceleration, with positions in 5G MEC and edge cloud; Xinqiyuan is pushing its self-developed chips into cloud, 5G/6G, finance and transport while accelerating its second-generation DPU. The opportunity, Wang argues, is to use the DPU as a lane-change moment: build domestic brands, participate in setting industry standards, and develop an autonomous industry chain – the new generation of DPU chips has a real chance to break the international monopoly in computing chips.
Want to talk through what you just read?
We work end to end, from product design to launch. Tell us about your scenario and we can start there.
Shenzhen XZS Technology Co., Ltd. · Tel 0755-82494862 · 13760424861