Tired of GPU/CPU? A new kind of computing power chip appeared! Big factories and capital are popular!

At the 2020 GTC conference, NVIDIA founder and CEO Huang Renxun said that DPU has become an important part of modern and secure accelerated data centers. The combination of CPU, GPU, and DPU can form a fully programmable single AI computing unit, providing unprecedented security and computing power.

DPU, the full name of is Data Processing Unit, is the third important computing power chip in data center scenarios after inheriting CPU and GPU, providing computing engines for high-bandwidth, low-latency, and data-intensive computing scenarios.

There are already many DPU players at home and abroad. According to the analysis of THENEXTPLATFORM, in 2020, the major foreign players include Broadcom, Intel, Nvidia, Netronome, Pensando, Fungible, and Xilinx, as well as the three giants of cloud providers.

Domestic manufacturers of DPU chip manufacturers mainly include Xinqiyuan, Dayu Zhixin, Zhongke Yushu, Nebula Zhilian, etc. In recent years, these manufacturers have been actively promoting and optimizing their own products, and capital is also actively investing in this field. For a while, Xinqiyuan and Nebulas Zhilian both announced the completion of different amounts of financing.

So what exactly is the DPU that has emerged in recent years and what functions does it have? What are the achievements of the major players, and what are the future prospects of the field?

What is DPU?

According to an article previously published by Nvidia, DPU is a new type of programmable processor that integrates three key elements:

The first is an industry-standard, high-performance, and software-programmable multi-core CPU, which is usually based on the widely used Arm architecture and closely cooperates with its SOC components;

The second is a high-performance network interface that can parse and process data at wire speed or the speed available in the network, and efficiently transmit data to the GPU and CPU.

The third is a variety of flexible and programmable acceleration engines that can offload applications such as AI, machine learning, security, telecommunications, and storage, and improve performance.

All these DPU functions are essential for realizing secure, bare-performance, cloud-native next-generation large-scale computing on the cloud.

DPU can be used as a standalone embedded processor, but it is usually integrated into SmartNIC (a network card that is a key component in next-generation servers).

DPU focuses on data processing, and the network data path acceleration engine needs to have at least the following 10 functions:

1. Parse, match and process data packets like OVS (Open Virtual Switch)

2. RDMA data transmission acceleration based on Zero Touch RoCE

3. Bypass the CPU through the GPU-Direct accelerator, and transfer data from storage and other GPUs directly to the GPU through the network

4. TCP communication acceleration, including RSS, LRO, checksum, and other operations

5. VXLAN and Geneve Overlay offloading and VTEP offloading for network virtualization

6. "Packet Pacing" traffic shaping acceleration for multimedia streaming, CDN (content delivery network), and new 4K/8K IP video (such as RiverMax based on ST 2110 specification)

7. Telecom Cloud RAN's precise clock accelerator, such as 5T for 5G (precise clock scheduling 5G wireless message transmission technology) function

8. Online IPSEC and TLS encryption acceleration, but does not affect other running acceleration operations

9. Support virtualization such as SR-IOV, VirtIO, and PV (Para-Virtualization)

10. Security isolation: such as the root of trust, secure boot, secure firmware upgrade, and lifecycle management of containers and applications based on identity authentication, etc.

Which domestic manufacturers are deploying DPU?

Datacenter construction drives the future development of DPU. According to Fungible and Nvidia's predictions, the DPU used in data centers will reach the same level as data center servers. At present, the number of new servers is about tens of millions each year. Each server may not have a GPU, but there must be one or more DPUs. If each DPU is calculated at 10,000 yuan, this will be a market size of hundreds of billions.

Faced with such a large market scale, and because it is an emerging product type, the gap between domestic and foreign companies is actually not very large. At present, many domestic manufacturers are accelerating their deployment, and new developments are reported from time to time. At the same time, chips are a huge capital For industries that invest in research and development, we can see that capital is also actively investing, and together with enterprises, accelerate the promotion of large-scale mass production applications in this field.


Xinqiyuan is a company that provides core chips and systems for ultra-large-scale telecommunications and enterprise-level intelligent networks. It is committed to building a DPU industry ecosystem, quickly deploying 5G cloud data centers and high-end EDA markets, and providing cloud data centers with the best intelligence The network card provides the most competitive emulation EDA solution for chip design companies.

Xinqiyuan has a large number of core products. The companys smart network card is currently the only complete solution for mature DPU based on SoC architecture in China. It has independent intellectual property rights and has matured mass production. It can provide chips, boards, and drivers. Software and a full set of cloud network solution products.

According to reports, the first batch of smart network card orders from China Mobile Suzhou Research Institute is the products purchased by Schiphol. At the network level of China Mobile's big cloud system, Schiphol's smart network card transparently uninstalls the key cloud network flow table of the SDN controller to the network card. Hardware acceleration can completely release a large number of server CPU and memory resources consumed by network services, and comprehensively improve the efficiency of China Mobile's cloud machine virtualization. The two parties have established a joint laboratory to jointly promote the development of next-generation smart network cards.

In February of this year and June of this year, Schiphol completed Pre A2 and Pre A3 rounds of financing of several hundred million yuan. The Pre A3 round of financing was jointly invested by SIG Hainer Asia, Pudong Science and Technology, Jingchen Semiconductor, and Yimei Investment. , The company said that this round of financing will continue to be used to attract global cutting-edge R&D and management talents to join the core source team, and to start the next stage of DPU chip technology research and development and market expansion.

Dayu Zhixin

Dayu Zhixin is committed to DPU product design, R&D, and service. The core team has more than ten years of experience in cloud computing platform design, R&D, and operation. It has served Baidu, Alibaba, Meituan, Intel, Cisco, and other companies. The use of the platform and the construction of the infrastructure have a clear understanding and rich practical experience.

Dayu Zhixins DPU products have high-speed network processing, seamless connection with distributed storage, and bare-metal management. They can be widely used in cloud computing, data centers, edge computing, 5G, and other fields. The products are oriented to enterprise customers. The testing of the first-generation product has been completed and is in the process of mass production. The product is expected to be shipped in the third quarter of this year. The first-generation products are mainly used in edge platforms, traffic orchestration, and management and control, etc., and have begun to develop 100G products, which will face a broader cloud computing scene.

The company announced on the 9th of this month that it has completed a Pre-A round of financing of tens of millions of yuan, invested by Huayi Venture Capital and Qiji Innovation. According to reports, this round of financing will be used for DPU product research and development, production investment, and high-end talents In the early days of its establishment, Dayu Zhixin also received an angel round of financing jointly invested by Zhongke Chuangxing and Unique Capital.

Zhongke Yushu

Founded in 2018, Zhongke Yushu focuses on the research and development of special processors, and is committed to solving the core problem of insufficient general computing power in the post Moore's Law era, and providing core chips and solutions for the intelligent era.

The company innovatively proposed software-defined accelerator technology, independently developed the KPU chip architecture, and created the industry's first DPU chip and smart network card series products and solutions with integrated network database acceleration functions.

In 2019, Zhongke Yushu taped out the K1 chip based on the self-developed KPU architecture. The current solutions are mainly used in financial scenarios. On March 31 this year, the company announced its next-generation DPU chip plan. Based on the self-developed KPU chip architecture, the company will launch a new-generation DPU chip K2, which can be used in high-end technologies such as financial technology, data centers, hybrid clouds, and edge computing. Bandwidth, low latency, and data-intensive computing scenarios.

K2's functions include complete L2/L3/L4 layer network protocol processing, which can process up to 200G network bandwidth data. It integrates database and big data processing capabilities and directly faces OLAP, OLTP, and big data processing platforms. It also includes machine learning computing cores and secure encryption computing cores. The chip is expected to be released by the end of 2021.

Nebulas Connect

Nebulas Zhilian focuses on the basic interconnection communication architecture of the data center and the research and development of DPU chips and is committed to building an intelligent connection and open ecosystem for the computing power of the digital world, making cloud computing and data centers a solid foundation for building a future digital society.

The company has gathered top experts in the ICT field from Silicon Valley, Israel, Canada, and other places. This team that has successfully achieved the development of communication and network chips in multiple fields in the past is the only domestic team that has reached the global leading level and is truly capable of developing DPU and The team and core members of its ecosystem have participated in the entire process of network big chip exploration, development, optimization, and transcendence. They have system-wide and long-cycle engineering development experience and have experience in cloud computing, the Internet, and operators. In-depth understanding of scenarios and the ability to design products that meet different market needs.

Nebulas Zhilian recently announced that it has completed a pre-A round of financing of several hundred million yuan. This round of financing is led by the old shareholder CDH VGC. The existing investors Hillhouse Ventures and Walden International continue to invest in; BAI Capital, Fosun, Huajin Investment, Gimpo Investment, Jiayu Fund, Songhe Capital, Wofu Capital, and many other investment institutions followed in the investment. The company only completed the angel round of financing three months ago.

Nebulas Zhilian believes that the DPU track meets the conditions for the growth of independent chip companies, regardless of the dispersion of the market or the reproducibility of the market segments. After completing this round of financing, Nebulas will continue to introduce research talents in multiple fields to build technical in-depth competitiveness, and introduce open source and ecological construction and operation talents to build ecological influence and new business models.

What is the future prospect of the DPU chip track?

Judging from the above content, there are not many manufacturers in the DPU chip layout, because it is an emerging track after all, but from the financing situation of several domestic companies, the degree of capital's favor for DPU is very high, you can see At that time, Xinqiyuan, Dayu Zhixin, and Nebulas Zhilian have all recently announced the completion of financing. Among them, Xinqiyuan and Nebulas Zhilian have completed three rounds of financing within two to three months. Why is capital optimistic about this new track?

First of all, judging from the possible future market size of this industry, it can be said to be quite huge. The current global demand for computing power has long surpassed the growth rate of computing power. The development of cloud data centers is facing disruptive changes in software and hardware, and DPU has emerged at the historic moment. , Responsible for data center transmission and data processing, and together with CPU and GPU constitute the three pillars of future computing.

DPU can free up the computing power of the CPU, release the load of the server, and significantly reduce the overall cost with low power consumption. In the future, each server must require one or more DPUs, just like each server must be equipped with a network card, just like the above As mentioned, the DPU used in the data center will reach the same level as the data center server, and the current server growth is about tens of millions each year, and the market size of the DPU reaches 100 billion.

In addition, in addition to the huge market size, downstream data center manufacturers are already actively deploying DPU products, including China Mobile, Amazon, Alibaba, and other giant manufacturers. In the future, driven by the demand of these downstream manufacturers, the products and market of DPU chips It is also easier to mature and scale.

On the whole, the future prospects for DPU applications in data centers are promising, but there are also areas worth paying attention to. For example, although large cloud service vendors have deployed a large number of DPU products, the real commercialization of DPU products is not yet successful, because the current market DPU product function coverage and scenario coverage are insufficient, and it is difficult to meet the needs of different customers for the rapid use of DPU products. Therefore, in the future, DPU chip manufacturers, cloud service providers, and capital will need to work together to actively optimize products and promote applications to jointly promote the development of the DPU industry.