Original link: https://www.latepost.com/news/dj_detail?id=1193
A new type of chip that has been around for less than a decade, DPU (Data Processing Unit, data processor), has now attracted the investment of many large Internet companies. They either build it themselves, such as Amazon and Alibaba, or invest in startups, such as Tencent, Meituan, ByteDance and Baidu.
“LatePost” exclusively learned that Tencent recently invested in a new round of financing of hundreds of millions of yuan in DPU startup Yunbao Smart. Other investors include Temasek, Shenzhen Venture Capital, Yicun Capital, an investment platform under Huaxi Village, and Minsheng Bank’s listing finance. Holding platforms CMBC Capital and Yaotu Capital, Yunbao’s post-investment valuation is about 9 billion yuan.
This is the third time Tencent has invested in this DPU company, which has been established for less than 2 years. According to industry and commerce information, Tencent currently holds 26.31% of Yunbao’s shares, making it its largest external shareholder. Most of Tencent’s stakes in investee companies do not exceed 20%.
Cloud Leopard was founded at the end of 2020 by Xiao Qiyang, general manager of Broadcom’s pre-processor and wireless infrastructure business unit in Greater China. Xiao Qiyang is not a typical chip executive, but an experienced entrepreneur. In 2002, he founded RMI in Silicon Valley, and its chip products are used in network communication equipment such as routers and base stations; in 2009, RMI was acquired by NetLogic, a US listed company, and NetLogic was acquired by Broadcom for $3.7 billion two years later. Xiao Qiyang then joined Broadcom until 2020 to start the creation of Cloud Leopard.
Cloud Leopard’s core members come from companies such as Broadcom, Intel, HiSilicon and Arm, and the team currently has nearly 400 people. According to “LatePost”, the cloud leopard’s first DPU SoC (system-on-chip) is in the development stage and is planned to be mass-produced next year. Cloud Leopard released the FPGA-based version of the DPU Yunxiao at the end of last year.
DPU is now attracting attention because it constitutes a new generation of pillar chips in the data center together with CPU and GPU/AI chips, and is considered to be a big opportunity to compete with CPU and GPU. Its main customers are large cloud computing companies and Internet companies. According to the forecast of Canalys, a research institute, by 2023, the size of China’s DPU market will be close to 20 billion yuan.
In addition to DPU, Tencent participated in the 2 billion A round of financing of GPU startup Moore Thread at the end of last year. Previously, Tencent also participated in 4 rounds of financing of AI chip company Suiyuan Technology, and currently holds 20.5% of the shares in Suiyuan. Suiyuan’s products are also big chips serving cloud data centers.
The birth of DPU stems from the gradual “unloading” process of data center CPUs.
In a traditional data center, the CPU has several roles: not only to handle various business tasks and general computing, but also to be responsible for the operation and maintenance of the entire server cluster, and to realize functions such as network, storage, security, and virtualization. This is like a store clerk, who is not only responsible for direct business work such as selling goods, but also moving things, arranging goods, and managing logistics… This consumes a lot of CPU resources, so that the CPU cannot focus on business tasks.
After 2012, the GPU, which was originally used for graphics rendering, and the AI chip specially developed for AI acceleration gradually took away part of the general-purpose computing work of the CPU. They are better at handling the rapidly growing AI computing demands.
Then in 2014, Amazon began developing a prototype of the DPU with Israeli startup Annapurna Labs. It can be understood as an “enhanced version” of the smart network card that realizes server network interconnection. The DPU is responsible for allocating different data processing tasks to the CPU or GPU reasonably; it further takes over the network, security, virtualization, storage and other data center operation and maintenance tasks from the CPU, so that the CPU can focus on processing business tasks and reduce computing resources. waste and improve overall efficiency.
After working with Annapurna Labs for a year, Amazon bought the company for $350 million in 2015 and launched Nitro in 2017, which is considered a sign of the emergence of DPUs. At the NVIDIA Technology Conference in 2020, NVIDIA founder Huang Renxun proposed that CPU+GPU+DPU would be a system solution for data centers; in the same year, NVIDIA acquired Mellanox, a leading global smart network card company, for $6.9 billion, aiming to develop its own DPU products. , this transaction drove the DPU craze.
Intel launched IPU (Infrastructure Processing Unit, similar in function and concept to DPU) in June last year; less than a month later, Marvell released its own DPU; in April this year, AMD acquired the US DPU established in 2017 for $1.9 billion Startup Pensando.
In China, Alibaba and Huawei choose to develop their own DPUs like Amazon. Ali released the virtualization technology Shenlong architecture in 2017, the purpose is to reduce the consumption of CPU resources by virtualization, and its hardware part can realize functions similar to DPU. The Hi1822 network card released by Huawei HiSilicon in 2020 also has some DPU functions, and it has been applied to Huawei’s self-developed TaiShan server.
Tencent, Meituan, Byte and other major data center chip customers choose to invest in start-ups.
Since 2020, a number of DPU companies such as Cloud Leopard, Yunmai Xinlian, Nebula Zhilian, Dayu Zhixin, and Zhongke Yushu have been born in China, and some earlier companies such as Xinqiyuan have turned to DPU in the past two years. direction.
Tencent has continuously invested in Yunbao, Byte has invested in Yunmai Xinlian last year (established in May 2021), and Meituan and Baidu have invested in Nebula Zhilian (established in March 2021) last year and this year respectively. Yunmai Xinlian was founded by Liu Yongfeng, who had worked at Cisco, Alibaba and Huawei, and Nebula Zhilian was founded by Xia Lusheng, who had worked at Huawei for many years.
It is also a large chip in the data center. Compared with the very general CPU and GPU, DPU needs to be more closely integrated with customer business scenarios.
An investor in a DPU company explained that different customers have different corresponding data types and data structures due to different business scenarios. At the data processing level, this is manifested by differences in packet size, task priority and performance focus— —For example, data processing related to live broadcast and games focuses on low latency, and data processing of financial services focuses on security. The infrastructure architecture of different customer data centers, that is, the mix of servers, chips and systems is also different, which also requires some targeted adaptation of the DPU.
“Without in-depth cooperation with major customers, it is difficult for DPU to achieve its original intention of improving efficiency,” said the investor.
A DPU practitioner believes that DPU can achieve good versatility, the premise is that you have a deep understanding of various business scenarios, know how to disassemble the scenarios into corresponding data processing processes, and screen out the most mainstream and most commercially valuable scenarios. . “DPU companies can’t support all business scenarios endlessly, but to hit 90% of the common needs at the cheapest cost.” He said. Cooperating with major customers with a large amount of chips and rich scenarios can help DPU companies accumulate business understanding.
This is also why DPU startups are willing to accept investment from a large Internet company at a very early stage and allow it to take a large share. They prefer to establish relationships with potential customers in advance, rather than pursuing the status of third-party neutral suppliers.
Tencent is one of the earliest investors of Cloud Leopard, and it may also be the first batch of customers of Cloud Leopard. According to “LatePost”, Tencent is already testing some chip functions of the clouded leopard.
Another feature of the DPU is that it is a new chip type. Here’s where the opportunity lies: There are no monopoly giants in this market yet. Globally, only Amazon’s self-developed Nitro has been used on a large scale. According to a blog published by Amazon AWS senior vice president James Hamilton in 2019, at that time, Amazon’s Nitro chip consumption had reached millions a year, and Nitro was not sold. The DPU products of Nvidia, Intel, AMD and other companies are still in development or iteration.
This also brings another risk that is different from CPU and GPU entrepreneurship: the entire DPU industry is at an early stage, the product form is uncertain, and the company may choose the wrong route.
In terms of chip form, DPU currently has two forms based on SoC (system-on-chip, DPU’s SoC generally includes CPU and ASIC, that is, dedicated integrated chip) and FPGA-based. The former has low power consumption, high performance, and low cost; the latter is more flexible and easier to respond to business changes, but at a higher cost. Amazon’s Nitro uses ASIC architecture, and Ali’s Dragon is currently in FPGA form.
A DPU practitioner believes that SoCs containing ASIC will be the mainstream of future DPUs. Compared with FPGA solutions, its cost is reduced by 50% and its performance is improved several times. There is also an industry view that due to different customer needs, the two forms will coexist for a long time.
In terms of function, an investor in the chip field told LatePost that DPU can be divided into “CPU faction” and “network faction”. The former DPU will contain a CPU with strong processing power, usually an Arm IP CPU, which can handle most of the data center operation and maintenance tasks related to data scheduling, storage, etc. Nvidia, Alibaba, and Yunbao are all in this line. Another route emphasizes network capabilities. Amazon’s Nitro, Intel’s IPU, AMD’s Pensando, and Chinese startup Yunmai Xinlian are more focused on this route.
It is also because the DPU is still in development and the shape has not been fully determined, and the boundary between it and its “predecessor” smart network card is blurred. An investor told “LatePost” that many companies are working on smart network cards under the banner of DPU, and have quickly gained a lot of investment and high valuations, but have not developed leading products.
The most intuitive and effective way to distinguish is to look at customer feedback: no matter what the architecture is or how it is implemented, a competitive DPU product should significantly improve data center efficiency.
It’s not yet time to get customer feedback. Like CPU, GPU and other large data center chips, DPU has a complex structure, long development cycle and large investment. A DPU company executive said that developing a mass-produced DPU SoC from 0 will cost at least $300 million and take about 2 years to develop. Most of China’s DPU startups started in 2020 and 2021. Including the early team formation time, the development cycle is mostly about 1 year, and the mass-produced products have not yet been launched.
When Chinese companies make GPUs and CPUs, it is difficult to compete with giants and break through ecological barriers; the challenge of making DPUs is that the entire track is very new and faces greater uncertainty, and there is no standard answer that can be learned.
This article is reprinted from: https://www.latepost.com/news/dj_detail?id=1193
This site is for inclusion only, and the copyright belongs to the original author.