The chip company Nvidia is working hard to make DPU (Data Processing Unit, data processor) the third largest computing pillar after CPU and GPU. Just as it simultaneously promoted the development of the CUDA ecosystem when developing GPUs, actively developing the DOCA ecosystem in the DPU era is similar to NVIDIA.

DPU is regarded as the third largest chip after CPU and GPU, and is responsible for processing data tasks that "CPU can't do well, GPU can't do it". The global chip giants NVIDIA and Intel have come to an end one after another, and a series of startups such as Zhongke Yushu, Yunbao Intelligence, and Dayu Zhixin have also emerged in China. The downstream cloud computing leaders Amazon's AWS, Alibaba Cloud and other manufacturers have already laid out... DPU industry It is ushering in the competitive landscape of a hundred schools of thought contending.

Since the release of the first DPU product BlueField-2 DPU series in October 2020, NVIDIA has planned multiple generations of products, and plans to launch the BlueField-3 DPU with stronger performance in 2022 and the BlueField-4 integrated with its GPU module in 2024. DPU. And DOCA (Data Center Infrastructure On A Chip Architecture), namely "online data center infrastructure architecture", is a software development platform tailored for BlueField DPU, the main purpose is to create a comprehensive open development platform for developers The platform supports developers to perform simple and flexible software development on BlueField DPU, allowing developers to quickly create BlueField DPU and accelerate some high-performance applications and services.

In May of this year, NVIDIA released DOCA version 1.3. Compared with the previous three versions, it is already a relatively complete software stack. Developers can easily build and develop on the local BlueField DPU or X86 development container based on this software stack. environment to quickly develop the application or service you want to implement.

At a recent media conference, Nvidia introduced the latest developments in its network technology.

Software and hardware ecology is the soul of DPU

According to Cui Yan, an expert in NVIDIA network technology, DOCA is a software framework tailored for BlueField DPU. Its main purpose is to create a comprehensive and open development kit for developers, and to support the majority of developers to do simple tasks on BlueField DPU. Flexible software development allows developers to quickly create applications and related services on the BlueField DPU. "In fact, its positioning is the soul of DPU, which is equivalent to unlocking DPU-related functions with software."

According to CCID Consulting, the global DPU market size will exceed 10 billion US dollars from 2023, and enter the fast lane of growth with an annual growth rate of more than 50%. In China, the DPU market size will also exceed 30 billion yuan in 2023, and there will be a jumping growth. The widespread adoption of DPU will spawn a large number of new applications and services, driving the growth of related hardware and software system platforms.

For the advancement of a new technology, especially at the large-scale implementation level, the driving of application and ecology is crucial.

Jiwei.com learned that in China, NVIDIA mainly builds its own ecology through three aspects: DOCA China developer community, DOCA training courses and developer training camps, and NVIDIA partner network.

The DOCA Chinese developer community started preparations in April 2021, and has gradually established the domestic ecosystem of DOCA. Cui Yan said that the DOCA Chinese developer community is continuing to expand, and more than half of the world's registered developers are from China. In addition, in terms of building DOCA applications and industry solutions, bare metal cloud acceleration network platforms, high-performance distributed storage, digital twin infrastructure, and supercomputing network platforms have all made progress.

In terms of partners, NVIDIA is currently cooperating with manufacturers such as VMware, Palo Alto Networks and Juniper Networks to expand the application scenarios of NVIDIA BlueField DPU and DOCA software architecture in platforms, infrastructure, storage, network security, 5G and edge computing. s solution.

Application Drives DPU Development

The advent of DPU comes from application drivers. "Modern hyperscale clouds are driving new architectures in the data center," said Jen-Hsun Huang, founder and CEO of NVIDIA when NVIDIA announced its first DPU. "A new type of processor designed to handle data center infrastructure software is needed to offload and accelerate virtualization, Huge computing load for networking, storage, security, and other cloud-native AI services."

With the deep integration of artificial intelligence and data center tasks, it has become difficult for CPUs to support data center workloads alone, and building a new data center architecture has become an industry consensus. In the new architecture of the "three U" of CPU, GPU and DPU, DPU can provide security, acceleration and infrastructure for various application loads in various environments of cloud, data center and edge computing.

In recent years, global chip leaders including Nvidia, Intel, Marvell, Broadcom, and many startups, as well as major cloud service vendors including AWS and Alibaba Cloud, have been deploying their own data processors, which shows that everyone has also recognized practical applications. Problems encountered in the data center.

However, Jiwei.com noticed that at the large-scale application level, the only ones in the world that really realize the large-scale commercial DPU architecture are Amazon's AWS and Alibaba Cloud. Among them, AWS adopts an Arm core-based solution. In 2017, AWS officially launched Nitro, which offloads networking, storage and security tasks to dedicated devices based on the Arm architecture. The core MOC card of the X-Dragon system architecture proposed by Alibaba Cloud adopts the form of FPGA+CPU.

In this regard, Cui Yan explained that there are three perspectives for the large-scale application of DPU: one is the large-scale application brought by the accelerated computing model through the data center, the DPU is the business application load of the CPU and the artificial intelligence and machine learning workload of the GPU. Provide data transmission and data processing, performance-cost ratio, infrastructure versatility and software and hardware iterability become the focus elements; one is customized design for the customer's own data center infrastructure application scenarios, through large-scale deployment of DPU to meet specific application scenarios needs and solves problems in specific scenarios, but it brings challenges to customers’ development capabilities and resource investment; in addition, through the large-scale applications brought by the ecosystem, starting from the diverse application needs of customers, based on an open and standardized platform. Provides a universal integration solution for innovative applications, ready to integrate and deploy, but requires a high-quality ecosystem.