Following the AI chip, DPU (data processor) has become a new hot spot for semiconductors.
Global chip giants Nvidia and Intel have successively ended one after another. A series of startups such as Zhongke Yushu, Yunbao Intelligence, Dayu Zhixin, Yunmai Xinlian, Nebula Intelligence, and Edge Zhixin have emerged in China. The downstream cloud computing leaders AWS, Alibaba Cloud and other manufacturers have also long been deployed......... The DPU industry ushered in a period of controversy among hundreds of schools of thought.
On December 21, Zhongke Yushu announced that it had completed a round of A+ financing of hundreds of millions of yuan, led by Maxxing Investment and Kunlun Capital. This is the second financing of hundreds of millions of yuan this year after completing the A round of financing at the end of July; Yunmai Xinlian, established in May, received investments from IDG, ByteDance and Biren Technology in October.
DPU is also stirring up the secondary market. On the 21st, Zuojiang Technology responded to investors that the company's self-developed programmable network security processing chip has DPU-related characteristics requirements in function definition and architecture design, and its stock price has risen for three consecutive days.
How much opportunity does domestic DPU manufacturers have in this new arena? Who can stand out?
The next golden track?
In the data center era, the situation where CPU is dominant will never be returned.
DPU is regarded as the third large chip second only to CPU and GPU, and is responsible for handling data tasks such as "If the CPU is not done well, the GPU cannot do it." With the surround of manufacturers such as Nvidia, Intel's data center market will be constantly eroded. The concept of
DPU was first proposed by Silicon Valley startup Fungible. This concept became a hit after Nvidia acquired network solution manufacturer Mellanox and repackaged the defined DPU. At present, the market is still in its early stages of development, and the definitions of different manufacturers are different. The mainstream DPU definition is proposed by NVIDIA, that is, DPU is a general-purpose processor that integrates data center infrastructure on chips.
The "2021 China DPU Industry Development White Paper" released by CIDI Consulting divides DPU solutions into three types: First, it is based on a general multi-core processor and multi-core Arm as the core, winning with the masses, and has good programmability flexibility, but its application is not targeted enough, and its support for special algorithms and applications does not have much significant advantages compared with general CPUs; second, a heterogeneous core array based on dedicated cores has strong targeting and good performance, but sacrifices some flexibility; the third route combines the advantages of the first two, that is, combining the programmable flexibility of general processors with dedicated acceleration engines, which is becoming the latest product trend.
Since the current concept and technical standards of DPU are not unified, each company outputs corresponding solutions based on its own advantages, forming a competitive landscape of hundreds of schools of thought.
On the one hand, global chip leaders such as Nvidia, Intel, Marvell, Botong are actively making plans and setting their own standards; on the other hand, many startups have joined this blue ocean. At the same time, major cloud service manufacturers including AWS and Alibaba Cloud are all laying out their own data processors.
According to First Financial , AWS, the world's largest cloud service manufacturer, adopts a solution based on Arm core. In 2015, AWS acquired Annapurna Labs, a manufacturer that designed chips based on the Arm architecture for hyperscale data centers and edge computing ; in 2017, AWS officially launched Nitro to offload network, storage and security tasks on dedicated devices based on the Arm architecture.
, while the core MOC card of the X-Dragon system architecture proposed by Alibaba Cloud uses the form of FPGA+CPU. Some industry insiders said that Microsoft used a pure FPGA solution in the early stage and is currently moving towards FPGA+CPU.
In addition to self-development, cloud manufacturers also look for suitable solutions through investment and other means. For example, Tencent invests in Yunbao Intelligence, and ByteDance invests in Yunmai Xinlian.
Dayu Zhixin CEO Li Shuang told Caixin that the current mainstream DPU market solutions include architectures dominated by Arm cores, FPGA+CPU architectures, and SoC architectures. The first two have been deployed in batches in cloud computing manufacturers, and the overall SoC solution is the future trend. "Our own product route is also based on this.Compared with the first generation Arm core, the richness and processing performance of the FPGA+CPU solution are much better, but the FPGA consumes more power and is more expensive. Meanwhile, our SoC project has been launched. "
Taobao Research Institute predicts that the global DPU market size will reach US$13.57 billion in 2025, and the Chinese DPU market size will reach US$3.74 billion.
CEDI Consultant's data is more optimistic. The agency expects that the global DPU market will usher in explosive growth from 2023 to 2024. China will usher in explosive growth from 2022 to 2023. It is estimated that by 2025, the global DPU industry market size will exceed US$24.53 billion, and the Chinese DPU industry market size will exceed 56.59 billion yuan.
Some industry insiders believe that the number of DPUs will be the same as the number of servers. Each server requires one or even more DPU products. The generation of
DPUs comes from application-driven. The slowdown of Moore's Law has caused the marginal cost of general-purpose CPU performance growth to rise rapidly. Yan Guihai, CEO of Zhongke Yushu, said that before 2010, the annualized bandwidth of networks increased by about 30%, and in recent years it reached 45%; while the performance growth of CPUs decreased from 23% 10 years ago to 3% in recent years. When the ratio of network bandwidth growth rate and CPU performance growth rate reaches 10 times, CPUs can hardly directly cope with the growth rate of network bandwidth.
"The core change in this era is that the increase in CPU performance cannot keep up with the speed of data generation. Before 2018, if I bought a 24-core CPU, it would have 2 cores to process traffic and 22 cores to process upper-level applications; after 2018, the traffic would increase rapidly, and 16 cores would need to process traffic, and only 8 cores would be used to process other tasks, which would be almost 2/3 of them would be scrapped. "Live Zhixin chief architect Li Fu told First Financial News.
In large data centers, traffic processing accounts for 30% of the computing resources. AWS calls these computing resources that have not yet run business programs and have to be taken up before they can access network data first. Datacenter Tax)”. The most direct role of
DPU is to act as the uninstall engine of the CPU, take over infrastructure services such as network virtualization and hardware resource pooling, thereby releasing valuable CPU resources to run upper-level applications, realizing the separation of data-centric business and infrastructure operations.
chip giant compete for the dominance of
October 2020, after acquiring Mellanox, Nvidia released its first DPU product - BlueField2 series. Nvidia founder and CEO O Huang Renxun said: "Modern hyper-large-scale cloud is promoting the new architecture of data centers. There is a need for a new processor designed to handle data center infrastructure software to offload and accelerate the huge computing load of virtualization, networking, storage, security and other cloud-native AI services. The era of BlueField DPU has arrived. "
Nvidia has planned multiple generations of products, planning to launch the stronger BlueField3 in 2022 and BlueField4 in 2024, which integrates its GPU modules.
Strategically, Nvidia acquires Mellanox to launch DPUs and tries to acquire Arm. Coupled with its traditional strong GPU, it will erode Intel's market share , directly threatening the latter's dominance in data centers.
's performance in the stock market is more direct. In July 2020, Nvidia City The value exceeds Intel and becomes the third largest semiconductor company in the world; more than a year later, Nvidia has become the world's highest-capital semiconductor company. As of December 28, Beijing time, Nvidia's market value was US$773.6 billion, almost four times that of Intel (market value of US$211.2 billion).
However, it remains to be seen whether Nvidia can successfully copy GPU success. Many industry insiders told First Financial that the recognition of Mellanox products in cloud service manufacturers was not high before. "Its disadvantage is that it really doesn't understand (cloud), including customers' tests and designs of its products. Before Mellanox was acquired, we tested many functions, and some changes in functions were also mentioned. ”
Many people believe that Nvidia's role in educating the DPU market and promoting concepts is greater than the practical application of its products.
Under Nvidia's step-by-step pressure, Intel is naturally not willing to lag behind. In June 2021, Intel released IPU (Infrastructure Processing Unit), benchmarking Nvidia's DPU, and launched two solutions based on FPGA and ASIC.
Guido Appenzeller, chief technology officer of Intel's data platform business unit, said that in the United States and China, the six major cloud service manufacturers currently use FPGA-based IPUs. "As bandwidth gets higher, we see them slowly turning to dedicated ASIC IPUs. Because there are many patented protocols, there is no rapid transition."
By using an IPU, the overhead associated with running infrastructure tasks can be offloaded from the server. The software of the cloud service vendor runs on the IPU itself, while the application of the tenant runs on the server CPU . This not only frees up resources on the server, but also optimizes overall performance, but also provides a separate and secure control point for cloud vendors.
It is worth mentioning that the CPU core used by the ASIC version of IPU is ARM architecture, not Intel's own X86 architecture , which means that Intel's voice is weakening, and competitors' architecture is selected to cater to customer needs.
First Financial News learned that Intel ASIC version IPU mainly cooperates with Google , Google put forward corresponding needs, and Intel "customized" it.
Li Fu believes that IPU focuses on the management of multiple CPUs, strives to integrate IaaS modules, reduce cluster management costs, and is more like a coprocessor. "Small cores similar to CPUs have become dual-channel CPU architectures in dual-channel servers. In order to meet different customer needs, Intel may adopt a super heterogeneous computing model of FPGA + small CPU + large CPUs to complicate the DPU/IPU market. We launched XPU chips based on PCIe Switch bus switching technology, which is a general data distribution chip prepared for this super heterogeneous computing architecture."
In addition, Xilinx , Marvell, Broadcom and other veteran chip manufacturers are also launching solutions for data centers.
Domestic startups are blooming everywhere
Domestic startups have also launched corresponding solutions one after another. Zhongke Yushu independently developed the KPU chip architecture, which is based on financial computing, data centers, cloud native, 5G edge computing and other scenarios. The second-generation DPU chip K2, which the company is developing, has completed design and verification work and is expected to be put into production in the first quarter of 2022.
On December 4, Dayu Zhixin’s first-generation DPU product Paratus V1.0 was officially unveiled. According to the company's plan, in 2021, Dayu Zhixin will complete the research and development of the first generation of products, customer testing and commercial delivery; by the end of 2021, the second generation of products will be successfully developed and the corresponding commercial contracts will be signed; by the end of 2021, Dayu Zhixin will start the research and development of its own chip (SoC solution).
On December 15, Yunbao Intelligent released a full-function Yunxiao DPU network card, which fully supports the integration of bare metal, virtual machine and container service resources and performance acceleration, mainly based on FPGA solutions. It is reported that the next generation of software-defined DPU chips are also under development as planned.
focuses on the edge of the PCIe technology stack. Intelligent core provides another perspective. After the data enters the XPU of its designed from the network card, it is distributed to the CPU, GPU, and SSD. According to its architecture, in the XPU-centric architecture, network card does not require intelligence, but only needs to provide basic network protocol functions, which is suitable for low-cost and low-latency application scenarios.
is different from CPU and GPU. DPU is currently in its early stages in terms of technical standards and ecology, and has not shown an absolute leader. Many industry insiders interviewed believe that China has rich data assets and a strong Internet industry. In this track, Chinese chip manufacturers may be expected to stand out.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-