Nvidia’s “carrying capacity” in the data center is not generally strong.
In April 2020, Nvidia announced the completion of a $6.9 billion acquisition of Israeli network chip maker Mellanox, pushing DPU Chips into the focus of the semiconductor field. Half a year later, NVIDIA launched the latest result of this acquisition – a new generation of DPU.
DPU, the full name of Data Processing Units (Data Processing Units), is called by NVIDIA founder Huang Renxun as the “three pillars of future computing” together with CPU and GPU. This new computing unit is like a dark horse in the data center, which has become very popular for a while, and has set off a new wave of investment both at home and abroad.
This year alone, domestic DPU startups Zhongke Yushu, Nebula Zhilian, Dayu Zhixin, Xinqiyuan, and Yunbao Smart have all announced the completion of new financing, many of which amounted to hundreds of millions of yuan.
In foreign countries, star DPU startups Pensando and Fungible have become the new favorites of data centers with strong “gold-absorbing” capabilities. Marvell founders Dai Weili and Zhou Xiuwen also founded a new DPU company, Dream Big Semiconductor, in 2019. Chip giants such as Intel, Marvell, and Broadcom are making early bets.
Does this increasingly popular new track really have the potential to share the data center computing market with CPU and GPU?
▲Domestic DPU startups to announce new financing in 2021
01.
Do the computing tasks that the CPU can’t do well and the GPU can’t do
“Each server may not have a GPU, but it definitely has a DPU, just like every server must be equipped with a network card.” Yan Guihai, founder and CEO of Zhongke Yushu, is very optimistic about the future of DPU, and believes that the emergence of DPU is one of heterogeneous computing. Stage sign.
In recent years, with the explosion of massive data, computing tasks have become more and more intensive, and the traditional CPU-centric data center architecture is increasingly facing the bottleneck of performance improvement. Computing resources perform their duties.
Huang Renxun, founder and CEO of NVIDIA, once explained that modern data centers are defined by software, and traditional infrastructure running data centers consumes 20-30% of CPU core resources, and DPU is a new type of dedicated CPU that “reduces the burden”. processor.
▲The gap between CPU main frequency computing power and network data transmission rate upgrade (Source: Toubao Research Institute)
DPU mainly undertakes network, storage, and security acceleration processing tasks, and is designed to meet the dedicated computing needs of the network side, especially in scenarios with a large number of servers and strict data transmission rate requirements.
It takes over the data processing tasks such as network protocol processing, data encryption and decryption, and data compression, which the CPU is not good at, and takes into account the needs of transmission and calculation through the combination of the processor and the network card. The DPU is also seen as an evolutionary enhanced version of SmartNIC (SmartNIC) technology.
Nowadays, the data center advocates doing it in the form of a resource pool, and various resources are managed and expanded separately, and the DPU is like a connection hub, which can play the role of central scheduling management, and one end is connected to various CPUs, GPUs, SSDs, FPGA accelerator cards, etc. Local resources, one end is connected to network resources such as switches/routers.
“Everyone thinks that in the future data center cloud architecture, the DPU may replace the CPU and become the most critical and core chip.” said Wang Lin, a well-known investor in the semiconductor industry and a partner of Walden International. The data center is regarded as a large service, and the DPU is equivalent to the CPU in this large server.
In general, DPU can not only improve network transmission efficiency, but also free up CPU computing resources, thereby driving the cost reduction and efficiency increase of the overall data center.
02.
Cloud computing is just needed, and giants are in battle
For cloud computing manufacturers and large Internet companies, purchasing DPU products is an inevitable trend. Looking back, leading cloud computing manufacturers at home and abroad have taken the lead in trying similar ideas.
As early as 2013, Amazon Cloud Technology (AWS) began to study how to offload the tasks previously processed by the CPU, such as network and storage, to the NIC card, then acquired the Israeli chip maker Annapurna Labs, and released the typical product AWS Nitro in 2017.
This attempt by AWS not only frees up the computing resources of the CPU, but also reduces the cost of the data center. Alibaba Cloud in China has also applied a dedicated X-Dragon chip in the MOC card, the core component of its Shenlong server, to uniformly support the virtualization of network, I/O, storage and peripherals.
In the past few years, chip companies such as Broadcom, Mellanox, Xilinx, and Intel have all launched their own DPU products. In addition to strengthening their processing capabilities for network protocols and network security, they have also successively demonstrated many performances in network, storage, and computing. Advantage.
Different from GPU, DPU is more low-level and more diverse in application-oriented applications. Various manufacturers rushing in the front row of DPU also have different emphasis on specific functions. The general differences are as follows:
▲Comparison of major foreign DPU manufacturers
After the acquisition of Mellanox, Nvidia further enhanced its combination of DPU and GPU capabilities, and unveiled a new roadmap last year, claiming to increase DPU performance by 1,000 times in three years. Next year, NVIDIA will release the BlueField-3X DPU with AI capabilities.
▲ NVIDIA BlueField DPU three-year roadmap
This year, chip giants such as Intel and Marvell continue to add DPUs.
In June, Intel released a programmable network device IPU (Infrastructure Processing Unit, infrastructure processor), which is used to expand the functions of intelligent network cards, accelerate functions such as storage virtualization, network virtualization and security, and release CPU cores. Its first FPGA-based IPU platform has been deployed for multiple cloud service providers, and its first ASIC IPU is in testing.
In July, Marvell launched the OCTEON 10 series DPU using TSMC’s 5nm process technology, using the Arm Neoverse N2 CPU core for the first time and integrating a machine learning inference engine. The first OCTEON 10 products are expected to sample before the end of this year and are said to significantly increase the processing throughput of packets by a factor of 5 compared to the previous generation scalar processing engine.
▲Marvell OCTEON 10 architecture
Since DPU is in the early stage of development, the industry has not yet formed a unified standard and has not formed enough scale deployment, which has brought a good opportunity for domestic DPU startups.
03.
The market size will reach tens of billions
Domestic DPU startups become the darling of capital
At the SoC design technology forum during this year’s Nanjing World Semiconductor Conference, Wang Lin, a well-known investor in the semiconductor industry and a partner of Walden International, mentioned that DPU is the hottest concept in SoC entrepreneurship this year, and it has slowly started to grow since last year.
According to the forecast of Canalys Cloud Channels Analysis, by 2023, the size of China’s DPU market will reach 19 billion yuan. According to the forecast of the Head Leopard Research Institute, the DPU market size in China is expected to exceed 3.7 billion US dollars (about 24 billion yuan) in 2025.
▲By 2025, China’s DPU market size forecast (Source: Head Leopard Research Institute)
At present, China’s DPU products are mainly supplied by international chip giants such as NVIDIA, Intel, and Broadcom. At the same time, domestic DPU startups such as Zhongke Yushu, Dayu Zhixin, Xinqiyuan, Yisixin, Nebula Zhilian, and Yunbao Intelligent are forming to promote DPU A new force for domestic replacement.
▲Main DPU startups in China
From the point of view of the establishment time, the domestic DPU representative start-up companies were founded in the past six years, and the earliest establishment was Chip Qiyuan. Dayu Zhixin, Yisixin, and Yunbao Intelligent were founded in June-August 2020, while Nebula Zhilian was just established in March this year.
Zhongke Yushu was incubated from the Chinese Academy of Sciences and was established in April 2018. Based on its self-developed KPU (Kernel Processing Unit) chip architecture, Zhongke Yushu achieved the successful tapeout of the industry’s first database and time series data processing fusion acceleration chip in October 2019, and released a KPU-based architecture in May 2020. Field-specific accelerator card products.
At present, Zhongke Yushu has reached cooperation with many financial institutions such as Zhongtai Securities, Huatai Securities, China Securities Construction Investment, Jinzheng Co., Ltd., and Shanghai Stock Exchange Technology. In addition to the financial computing field, its DPU products can also be widely used in data centers, telecommunications, Internet of Things, Internet, hybrid cloud and other fields in the future.
▲ Zhongke Yushu DPU chip product line
Dayu Zhixin, also headquartered in Beijing, was established in June 2020. At the beginning of its establishment, it received angel round financing and was jointly invested by Zhongke Chuangxing and Weiyi Capital; in July this year, it announced the completion of tens of millions of yuan Pre-A The round of financing was invested by Huayi Venture Capital and Qiji Chuangtan.
Li Shuang, founder and CEO of Dayu Zhixin, once served as general manager of Meituan Cloud, director of Alibaba Network Technology, and chairman of Baidu System Technical Committee. The high-performance gateway lays the hardware foundation for the era of cloud computing in China. It has many years of experience in network equipment, server, processor customization and heterogeneous platform research and development.
It is reported that the members of the Dayu Zhixin team are from Alibaba Cloud and Google Cloud. They have personally led the implementation of cloud platforms and DPUs, and have complete DPU experience. Its first-generation smart NIC Dayu DPU series product, Paratus 1.0, claims to be the first commercial DPU product with a processing capacity of 25G in China. It has entered the production stage and is expected to be shipped in the third quarter of this year. The company has begun to develop 100G products.
Shanghai Xinqiyuan was founded in 2015, and this year, it has successively completed hundreds of millions of Pre-A2 and Pre-A3 rounds of financing. At present, Coreqiyuan focuses on 5G cloud data center and high-end EDA field, and has two core products: smart network card, SoC prototype and simulation system MIMIC.
Lu Sheng, founder of Chipqiyuan, graduated from Shanghai Jiaotong University with a bachelor’s degree and later obtained a master’s degree from Lamar University in the United States. He has held important positions in Broadcom, Marvell, ArtX and other companies successively. His core team has accumulated more than 20 years of R&D management experience in the chip field. .
It is reported that the chip Qiyuan smart network card is the only mature DPU complete solution based on SoC architecture in China.
▲Core Qiyuan intelligent network card
Shanghai Yisixin Technology was established in Shanghai in July 2020. Its founder, chairman and CEO Huang Yiren is an expert in the semiconductor chip industry and a leader in the DPU smart network card industry. It is reported that the technical team of Yisixin Technology is composed of core professionals and R&D personnel in the fields of network, switching and storage at home and abroad, and has profound technical strength in the fields of network, switching, storage and high-performance CPU.
In July this year, Essex Technology announced that it had delivered the first domestic self-developed commercial cloud-native smart network card Stargate-F1000-SN, with a power consumption of 30W to 40W for the whole card.
▲ Stargate-F1000-SN smart network card of Yisixin Technology
Shenzhen Yunbao Intelligent was founded by Dr. Sunny Siu Xiao Qiyang, co-founder of the former RMI company (later acquired by Netlogic/Broadcom) and industry elites. Its core team comes from well-known domestic and foreign companies such as Broadcom, Intel, Arm, HiSilicon, and Alibaba.
According to the official website information, Tencent, Sequoia Capital, Zhengxingu Capital, Yaotu Capital, SMIC Juyuan, ACE Equity Partners, Huaye Tiancheng, and OBN Capital are all investors of Yunbao Smart.
Before founding Cloud Leopard Intelligence, Xiao Qiyang served as the general manager of Broadcom’s processor and wireless infrastructure business unit in Greater China, the president and general manager of NetLogic’s Asia-Pacific region (acquired by Broadcom in 2012 for US$3.7 billion), RMI (acquired by NetLogic) acquisition) co-founder and president of Asia Pacific.
Zhuhai Nebula Zhilian was established in March this year. In April, it announced the completion of hundreds of millions of angel rounds of financing, and in July, it completed hundreds of millions of yuan in Pre-A round of financing.
It is reported that the team of Nebula Zhilian has successfully realized the development of multi-domain communication and network chips. The core members have participated in the whole process of exploration, development, optimization and surpassing of large network chips from 0 to 1, and have experience in the whole system and long-term engineering development. , has a deep understanding of cloud computing, Internet, operators and other scenarios, and the ability to design products to meet different market needs.
The DPU being developed by Nebulas will form an independent CAAS (communication service layer) between IAAS and PAAS, to achieve “many virtual one” physical resources and “one virtual multiple” nearly bare metal performance, simplify IAAS and improve resource utilization rate; offload the processing related to communication data flow in PAAS to improve the communication efficiency and performance of the application.
The main research products of other domestic startups can also be classified into the DPU category, but because the official website does not clearly mention the relationship with DPU, it is not included in this article. Readers are welcome to add more domestic companies that are developing DPU in the comment area. company.
04.
Conclusion: The curtain of DPU’s entrepreneurship has just opened
As one of the most important chips in the new data center, the DPU is accelerating its development. In the view of Yu Yong, CEO of Nebulas Zhilian, in the next ten years, the global cloud computing and data center infrastructure will usher in disruptive changes, and the DPU-based unified communication service platform (CAAS) is the core engine of this future architecture.
Judging from the current global DPU industry structure, product applications are still relatively scattered. Although international giants occupy technological and market advantages, the generation gap at home and abroad is not insurmountable. Fortunately, many domestic companies are already exploring the road to commercialization.
For the next development, Dayu Zhixin CEO Li Shuang judged that the DPU needs to work in two directions: one is to improve the virtualization efficiency, and the other is to improve the throughput of mobile data to help the data center network reach the 100G goal.
Although it will take time to realize the transformation of core technology into main income, or make the emerging concept of DPU accepted by more customers, but with the joint efforts of upstream and downstream manufacturers in the industry chain, I believe that there will be more companies in the future. Invest in this emerging track and jointly promote the prosperity of the domestic DPU ecosystem.
Nvidia’s “carrying capacity” in the data center is not generally strong.
In April 2020, Nvidia announced the completion of a $6.9 billion acquisition of Israeli network chip maker Mellanox, pushing DPU chips into the focus of the semiconductor field. Half a year later, NVIDIA launched the latest result of this acquisition – a new generation of DPU.
DPU, the full name of Data Processing Units (Data Processing Units), is called by NVIDIA founder Huang Renxun as the “three pillars of future computing” together with CPU and GPU. This new computing unit is like a dark horse in the data center, which has become very popular for a while, and has set off a new wave of investment both at home and abroad.
This year alone, domestic DPU startups Zhongke Yushu, Nebula Zhilian, Dayu Zhixin, Xinqiyuan, and Yunbao Smart have all announced the completion of new financing, many of which amounted to hundreds of millions of yuan.
In foreign countries, star DPU startups Pensando and Fungible have become the new favorites of data centers with strong “gold-absorbing” capabilities. Marvell founders Dai Weili and Zhou Xiuwen also founded a new DPU company, Dream Big Semiconductor, in 2019. Chip giants such as Intel, Marvell, and Broadcom are making early bets.
Does this increasingly popular new track really have the potential to share the data center computing market with CPU and GPU?
▲Domestic DPU startups to announce new financing in 2021
01.
Do the computing tasks that the CPU can’t do well and the GPU can’t do
“Each server may not have a GPU, but it definitely has a DPU, just like every server must be equipped with a network card.” Yan Guihai, founder and CEO of Zhongke Yushu, is very optimistic about the future of DPU, and believes that the emergence of DPU is one of heterogeneous computing. Stage sign.
In recent years, with the explosion of massive data, computing tasks have become more and more intensive, and the traditional CPU-centric data center architecture is increasingly facing the bottleneck of performance improvement. Computing resources perform their duties.
Huang Renxun, founder and CEO of NVIDIA, once explained that modern data centers are defined by software, and traditional infrastructure running data centers consumes 20-30% of CPU core resources, and DPU is a new type of dedicated CPU that “reduces the burden”. processor.
▲The gap between CPU main frequency computing power and network data transmission rate upgrade (Source: Toubao Research Institute)
DPU mainly undertakes network, storage, and security acceleration processing tasks, and is designed to meet the dedicated computing needs of the network side, especially in scenarios with a large number of servers and strict data transmission rate requirements.
It takes over the data processing tasks such as network protocol processing, data encryption and decryption, and data compression, which the CPU is not good at, and takes into account the needs of transmission and calculation through the combination of the processor and the network card. The DPU is also seen as an evolutionary enhanced version of SmartNIC (SmartNIC) technology.
Nowadays, the data center advocates doing it in the form of a resource pool, and various resources are managed and expanded separately, and the DPU is like a connection hub, which can play the role of central scheduling management, and one end is connected to various CPUs, GPUs, SSDs, FPGA accelerator cards, etc. Local resources, one end is connected to network resources such as switches/routers.
“Everyone thinks that in the future data center cloud architecture, the DPU may replace the CPU and become the most critical and core chip.” said Wang Lin, a well-known investor in the semiconductor industry and a partner of Walden International. The data center is regarded as a large service, and the DPU is equivalent to the CPU in this large server.
In general, DPU can not only improve network transmission efficiency, but also free up CPU computing resources, thereby driving the cost reduction and efficiency increase of the overall data center.
02.
Cloud computing is just needed, and giants are in battle
For cloud computing manufacturers and large Internet companies, purchasing DPU products is an inevitable trend. Looking back, leading cloud computing manufacturers at home and abroad have taken the lead in trying similar ideas.
As early as 2013, Amazon Cloud Technology (AWS) began to study how to offload the tasks previously processed by the CPU, such as network and storage, to the NIC card, then acquired the Israeli chip maker Annapurna Labs, and released the typical product AWS Nitro in 2017.
This attempt by AWS not only frees up the computing resources of the CPU, but also reduces the cost of the data center. Alibaba Cloud in China has also applied a dedicated X-Dragon chip in the MOC card, the core component of its Shenlong server, to uniformly support the virtualization of network, I/O, storage and peripherals.
In the past few years, chip companies such as Broadcom, Mellanox, Xilinx, and Intel have all launched their own DPU products. In addition to strengthening their processing capabilities for network protocols and network security, they have also successively demonstrated many performances in network, storage, and computing. Advantage.
Different from GPU, DPU is more low-level and more diverse in application-oriented applications. Various manufacturers rushing in the front row of DPU also have different emphasis on specific functions. The general differences are as follows:
▲Comparison of major foreign DPU manufacturers
After the acquisition of Mellanox, Nvidia further enhanced its combination of DPU and GPU capabilities, and unveiled a new roadmap last year, claiming to increase DPU performance by 1,000 times in three years. Next year, NVIDIA will release the BlueField-3X DPU with AI capabilities.
▲ NVIDIA BlueField DPU three-year roadmap
This year, chip giants such as Intel and Marvell continue to add DPUs.
In June, Intel released a programmable network device IPU (Infrastructure Processing Unit, infrastructure processor), which is used to expand the functions of intelligent network cards, accelerate functions such as storage virtualization, network virtualization and security, and release CPU cores. Its first FPGA-based IPU platform has been deployed for multiple cloud service providers, and its first ASIC IPU is in testing.
In July, Marvell launched the OCTEON 10 series DPU using TSMC’s 5nm process technology, using the Arm Neoverse N2 CPU core for the first time and integrating a machine learning inference engine. The first OCTEON 10 products are expected to sample before the end of this year and are said to significantly increase the processing throughput of packets by a factor of 5 compared to the previous generation scalar processing engine.
▲Marvell OCTEON 10 architecture
Since DPU is in the early stage of development, the industry has not yet formed a unified standard and has not formed enough scale deployment, which has brought a good opportunity for domestic DPU startups.
03.
The market size will reach tens of billions
Domestic DPU startups become the darling of capital
At the SoC design technology forum during this year’s Nanjing World Semiconductor Conference, Wang Lin, a well-known investor in the semiconductor industry and a partner of Walden International, mentioned that DPU is the hottest concept in SoC entrepreneurship this year, and it has slowly started to grow since last year.
According to the forecast of Canalys Cloud Channels Analysis, by 2023, the size of China’s DPU market will reach 19 billion yuan. According to the forecast of the Head Leopard Research Institute, the DPU market size in China is expected to exceed 3.7 billion US dollars (about 24 billion yuan) in 2025.
▲By 2025, China’s DPU market size forecast (Source: Head Leopard Research Institute)
At present, China’s DPU products are mainly supplied by international chip giants such as NVIDIA, Intel, and Broadcom. At the same time, domestic DPU startups such as Zhongke Yushu, Dayu Zhixin, Xinqiyuan, Yisixin, Nebula Zhilian, and Yunbao Intelligent are forming to promote DPU A new force for domestic replacement.
▲Main DPU startups in China
From the point of view of the establishment time, the domestic DPU representative start-up companies were founded in the past six years, and the earliest establishment was Chip Qiyuan. Dayu Zhixin, Yisixin, and Yunbao Intelligent were founded in June-August 2020, while Nebula Zhilian was just established in March this year.
Zhongke Yushu was incubated from the Chinese Academy of Sciences and was established in April 2018. Based on its self-developed KPU (Kernel Processing Unit) chip architecture, Zhongke Yushu achieved the successful tapeout of the industry’s first database and time series data processing fusion acceleration chip in October 2019, and released a KPU-based architecture in May 2020. Field-specific accelerator card products.
At present, Zhongke Yushu has reached cooperation with many financial institutions such as Zhongtai Securities, Huatai Securities, China Securities Construction Investment, Jinzheng Co., Ltd., and Shanghai Stock Exchange Technology. In addition to the financial computing field, its DPU products can also be widely used in data centers, telecommunications, Internet of Things, Internet, hybrid cloud and other fields in the future.
▲ Zhongke Yushu DPU chip product line
Dayu Zhixin, also headquartered in Beijing, was established in June 2020. At the beginning of its establishment, it received angel round financing and was jointly invested by Zhongke Chuangxing and Weiyi Capital; in July this year, it announced the completion of tens of millions of yuan Pre-A The round of financing was invested by Huayi Venture Capital and Qiji Chuangtan.
Li Shuang, founder and CEO of Dayu Zhixin, once served as general manager of Meituan Cloud, director of Alibaba Network Technology, and chairman of Baidu System Technical Committee. The high-performance gateway lays the hardware foundation for the era of cloud computing in China. It has many years of experience in network equipment, server, processor customization and heterogeneous platform research and development.
It is reported that the members of the Dayu Zhixin team are from Alibaba Cloud and Google Cloud. They have personally led the implementation of cloud platforms and DPUs, and have complete DPU experience. Its first-generation smart NIC Dayu DPU series product, Paratus 1.0, claims to be the first commercial DPU product with a processing capacity of 25G in China. It has entered the production stage and is expected to be shipped in the third quarter of this year. The company has begun to develop 100G products.
Shanghai Xinqiyuan was founded in 2015, and this year, it has successively completed hundreds of millions of Pre-A2 and Pre-A3 rounds of financing. At present, Coreqiyuan focuses on 5G cloud data center and high-end EDA field, and has two core products: smart network card, SoC prototype and simulation system MIMIC.
Lu Sheng, founder of Chipqiyuan, graduated from Shanghai Jiaotong University with a bachelor’s degree and later obtained a master’s degree from Lamar University in the United States. He has held important positions in Broadcom, Marvell, ArtX and other companies successively. His core team has accumulated more than 20 years of R&D management experience in the chip field. .
It is reported that the chip Qiyuan smart network card is the only mature DPU complete solution based on SoC architecture in China.
▲Core Qiyuan intelligent network card
Shanghai Yisixin Technology was established in Shanghai in July 2020. Its founder, chairman and CEO Huang Yiren is an expert in the semiconductor chip industry and a leader in the DPU smart network card industry. It is reported that the technical team of Yisixin Technology is composed of core professionals and R&D personnel in the fields of network, switching and storage at home and abroad, and has profound technical strength in the fields of network, switching, storage and high-performance CPU.
In July this year, Essex Technology announced that it had delivered the first domestic self-developed commercial cloud-native smart network card Stargate-F1000-SN, with a power consumption of 30W to 40W for the whole card.
▲ Stargate-F1000-SN smart network card of Yisixin Technology
Shenzhen Yunbao Intelligent was founded by Dr. Sunny Siu Xiao Qiyang, co-founder of the former RMI company (later acquired by Netlogic/Broadcom) and industry elites. Its core team comes from well-known domestic and foreign companies such as Broadcom, Intel, Arm, HiSilicon, and Alibaba.
According to the official website information, Tencent, Sequoia Capital, Zhengxingu Capital, Yaotu Capital, SMIC Juyuan, ACE Equity Partners, Huaye Tiancheng, and OBN Capital are all investors of Yunbao Smart.
Before founding Cloud Leopard Intelligence, Xiao Qiyang served as the general manager of Broadcom’s processor and wireless infrastructure business unit in Greater China, the president and general manager of NetLogic’s Asia-Pacific region (acquired by Broadcom in 2012 for US$3.7 billion), RMI (acquired by NetLogic) acquisition) co-founder and president of Asia Pacific.
Zhuhai Nebula Zhilian was established in March this year. In April, it announced the completion of hundreds of millions of angel rounds of financing, and in July, it completed hundreds of millions of yuan in Pre-A round of financing.
It is reported that the team of Nebula Zhilian has successfully realized the development of multi-domain communication and network chips. The core members have participated in the whole process of exploration, development, optimization and surpassing of large network chips from 0 to 1, and have experience in the whole system and long-term engineering development. , has a deep understanding of cloud computing, Internet, operators and other scenarios, and the ability to design products to meet different market needs.
The DPU being developed by Nebulas will form an independent CAAS (communication service layer) between IAAS and PAAS, to achieve “many virtual one” physical resources and “one virtual multiple” nearly bare metal performance, simplify IAAS and improve resource utilization rate; offload the processing related to communication data flow in PAAS to improve the communication efficiency and performance of the application.
The main research products of other domestic startups can also be classified into the DPU category, but because the official website does not clearly mention the relationship with DPU, it is not included in this article. Readers are welcome to add more domestic companies that are developing DPU in the comment area. company.
04.
Conclusion: The curtain of DPU’s entrepreneurship has just opened
As one of the most important chips in the new data center, the DPU is accelerating its development. In the view of Yu Yong, CEO of Nebulas Zhilian, in the next ten years, the global cloud computing and data center infrastructure will usher in disruptive changes, and the DPU-based unified communication service platform (CAAS) is the core engine of this future architecture.
Judging from the current global DPU industry structure, product applications are still relatively scattered. Although international giants occupy technological and market advantages, the generation gap at home and abroad is not insurmountable. Fortunately, many domestic companies are already exploring the road to commercialization.
For the next development, Dayu Zhixin CEO Li Shuang judged that the DPU needs to work in two directions: one is to improve the virtualization efficiency, and the other is to improve the throughput of mobile data to help the data center network reach the 100G goal.
Although it will take time to realize the transformation of core technology into main income, or make the emerging concept of DPU accepted by more customers, but with the joint efforts of upstream and downstream manufacturers in the industry chain, I believe that there will be more companies in the future. Invest in this emerging track and jointly promote the prosperity of the domestic DPU ecosystem.
The Links: SKM400GA123D XC6SLX9-2TQG144C