The largest DPU manufacturer in history (Part 1)

DPU It is a chip with data processing as the center ,2020 year NVIDIA Strategic China calls it CPU、GPU after “ The third main chip ”, Think “ It will become one of the three pillars of computing in the future ”. The cheetah Institute predicts , The data center area DPU It is about to be measured , With intelligent driving 、 The needs of other fields, such as the meta universe, are constantly being tapped ,DPU The penetration application field will continue to expand , China DPU The market size is expected to be 2026 reach 1095.3 One hundred million yuan .

DPU The hot market has attracted large domestic and foreign manufacturers to enter the market one after another , It has also spawned a number of start-ups . So what are the main DPU What about players ？

Foreign manufacturers

Nvidia

Nvidia Founded in 1993 year , Headquartered in Santa Clara, California, USA .1999 year ,Nvidia Defined GPU, It's a big push PC The development of game market , Redefined modern computer graphics technology .2020 year 4 month ,NVIDIA The official announcement has been made that Mellanox Acquisition , Product layout covers CPU、GPU and DPU.

NVIDIA BlueField DPU Bring innovation to modern data centers . Through a variety of advanced networks 、 Storage and security services are unloaded 、 Acceleration and isolation ,BlueField DPU Can be cloud 、 Various workloads in environments such as data centers or edge computing provide a secure and accelerated infrastructure .BlueField DPU Will be a powerful computing power 、 Complete on-chip infrastructure programmability and high-performance network , Support demanding workloads .

• Security from edge to center ：BlueField DPU A comprehensive security architecture that supports zero trust , Covering data center and edge computing .
• Provide elastic storage for expanding workloads ： With the help of NVMe over Fabric (NVMe-oF)、GPUDirect Storage 、 encryption 、 Elastic storage 、 Data integrity 、 Decompression and deduplication support ,BlueField Provides high-performance storage access solutions , Achieve ultra-low latency for remote storage comparable to direct attached storage .
• High performance and efficient network ：BlueField It is a powerful data center service accelerator , It can be used for both traditional and modern applications GPU Accelerated applications provide up to 400 Gb/s Ethernet or InfiniBand Connection speed , Release the host at the same time CPU kernel , To run applications other than infrastructure tasks .
• Software defined infrastructure ：NVIDIA DOCA Software Development Suite (SDK) Enable developers to leverage industry standards API Easily create high performance 、 Software definition 、 Cloud native DPU Accelerated Service .

NVIDIA BlueField-3 It is the first line speed processing software to define the network 、 Storage and network security 400Gb/s DPU.BlueField-3 Will be a powerful computing power 、 High speed network and extensive programmability are combined , Provide software defined hardware acceleration solutions for demanding workloads . From acceleration AI Calculation , To mixed clouds , Then to cloud native supercomputing and 5G Wireless network ,BlueField-3 Redefined the possibilities .

Official website ：https://www.nvidia.com/en-us/

AMD

AMD Semiconductor company was founded in 1969 year , Specifically for computers 、 The communications and consumer electronics industries design and manufacture a variety of CPU、GPU Such microprocessor .2022 year 2 month ,AMD Finally finished the right Xilinx Acquisition , The value is close to 500 The transaction of billion yuan is AMD brought Xilinx Of FPGA Programmable logic modules and related DSP engine 、AI Accelerator 、 Memory controller and other key technologies , by AMD The technical reserve has been replenished .

Xilinx Provided DPU/SmartNIC yes Alveo series ,Alveo Series based on FPGA, Can accelerate computing intensive applications , Including machine learning and reasoning 、 Data analysis 、 Video transcoding and many other workloads ,Alveo Performance ratio of the series CPU High performance 90 times , And it can be reprogrammed according to the specific requirements of users , Because the algorithm develops faster than the chip design cycle , Therefore, programmable hardware that can adapt to changing algorithms is needed .

Xilinx Alveo SN1000 It is the first in the industry to provide software defined hardware acceleration for all function uninstallation in a single platform SmartNIC.SN1000 SmartNIC Uninstall directly CPU Intensive tasks to optimize network performance , Its architecture can accelerate various custom uninstallations at line speed , Including support for customer build and third-party uninstall .SN1000 SmartNIC be based on Xilinx 16nm UltraScale+ framework , By low latency Xilinx XCU26 FPGA and 16 nucleus Arm Processor support .

2022 year 5 month ,AMD Announce the completion of Pensando Systems Acquisition , The transaction price is about 19 Billion dollars .Pensando Distributed service platform , It will pass through the high-performance data processing unit （DPU） And software stack extensions AMD Our data center Portfolio . These products have been sold at Goldman Sachs 、IBM Cloud、Microsoft Azure and Oracle Cloud Large scale deployment of enterprises in the cloud, etc .Pensando Of Elba SoC Is a focus on intelligent network switches DPU, Previous Capri DPU Be used for Aruba CX 10000 .

Official website ：https://www.amd.com/en

Intel

stay “Intel Vision 2022”, At the conference Intel Released its latest IPU The roadmap , Show me from 2022 - 2026 year IPU The overall planning of . Intel will continue ASIC + FPGA IPU Design , Its IPU The roadmap is as follows ：

• 2022 year ： Launched 200 Gbps IPU, code-named Mount Evans and Oak Springs Canyon.
• 2023/2024 year ： Introduction 400 Gbps IPU, code-named Mount Morgan and Hot Springs Canyon.
• 2025/2026 : Introduction 800 Gbps IPU.

Mount Evans yes Intel One of the first ASIC IPU, And Google Cloud Cooperative development , For high-end and very large-scale data center servers .Oak Springs Canyon yes Intel The second generation is based on FPGA Of IPU platform , The platform uses Intel Xeon-D and Agilex FPGA structure .

Intel IPU One of the key technologies is the fast programmable packet processing engine supported by all devices . Whether it's FPGA Or based on ASIC Products , Customers can use P4 Program it , And supports searching 、 change 、 Encryption and compression processes .

in addition ,Intel Also launched IPU Open source development kit for IPDK , It can be used for x86 Chips and Arm chip ( Such as Marvell Of Octeon) Write applications . The toolkit includes function blocks for customizing and defining workloads , This includes unloading package handling .（ More can be clicked ：IPDK： Open source development framework in the era of programmable infrastructure ）

Official website ：https://www.intel.com/

Marvell

‍

Marvell Founded on 1995 year , Headquartered in Silicon Valley , It has a research and development center in Shanghai, China , Is a global leading semiconductor manufacturer providing a full range of broadband communication and storage solutions .

Marvell Of OCTEON and ARMADA Devices are designed for wireless infrastructure and network devices , Including switches 、 Router 、 Security gateway 、 A firewall 、 Network monitoring and intelligent network card (SmartNIC), And support comprehensive and unified SDK And open source API, For a wide range of networks 、 Security and computing market applications .

Marvell Of OCTEON 10 DPU Series for very large-scale cloud workloads 、5G Wireless transmission 、5G RAN Intelligent controller (RIC) And marginal reasoning 、 Operator and enterprise data center applications and fanless network edge boxes are optimized .OCTEON 10 Use TSMC 5nm Process technology and ARM Of Neoverse N2 CPU kernel , Plus a generation OCTEON TX2 An array of functional building blocks , It also includes an engine that integrates machine learning reasoning 、 Inline encryption processor and vector packet processor IP And functions , And both can run in a virtualized way . As DPU An important addition to ,Marvell Also for OCTEON 10 Introduce internal machine learning (ML) engine .

• The first in the industry to adopt Arm Neoverse N2 Kernel 5nm DPU, With previous generations OCTEON comparison , Improved computing performance 3 times , Reduced power consumption 50%
• For inline ML/AI The innovative hardware accelerator offers more than software based reasoning 100 Times performance improvement
• be based on VPP The hardware accelerator has improved the packet processing speed 5 More than times
• Integrate 1 Terabit Switch 、 True inline encryption and highly programmable packet processing
• Data path support exceeds 400G
• Support the latest PCIe 5.0 I/O And DDR5 Memory

Official website ：https://www.marvell.com/

Broadcom

Broadcom Of Stingray Combined with powerful network controller 、 High performance ARM CPU、PCI Express 3.0、 Performance accelerators and DDR4 RAM, Take compute intensive applications from the host server's CPU uninstall .

Stingray It can provide high packet rate and low latency .Broadcom With NetXtreme E Based on the logic of the series controller , stay Stingray The core part of the system is designed NetXtreme-S BCM58800 chip , Then in the cluster configuration 8 The dominant frequency is 3 GHz Of Arm v8 A72 kernel . Besides ,Stingray It can also be configured 16 GB DDR4 Memory .

Broadcom It also uses TruFlow technology , This is a configurable stream Accelerator , Used to transfer common network flow processes to hardware . Judging from the published data ,TruFlow You can uninstall on your hardware such as Open vSwitch（OvS） Tasks like that . The company also claims that TruFlow In the hardware, many classic SDN Concept , Such as the classification 、 Matching and operation . therefore ,Stingray Equipped with two programmable components , namely TruFlow And by four 3 GHz Dual core Arm v8 A72 A cluster of complexes .

Official website ：https://www.broadcom.com/

Fungible

2019 year ,Fungible take DPU It is defined as a new type of data processing unit .Fungible Of F1 DPU It is the first in the industry 800Gbps Of DPU, It's also Fungible DPU The flagship product of the series .

On the architecture ,F1 DPU A large number of multi-core processors are integrated ,52 All the cores are the latest generation MIPS64 R6 kernel , It not only supports hardware virtualization, but also divides it into independent control units .F1 DPU The design of double launch pipeline is adopted , Equipped with 64KB Of L1 I-cache and 80KB Of L1 D-Cache, And L1 Caching supports data transfer between caches , Total on-chip L2 The cache reaches 32MB. Memory aspect ,F1 DPU In addition to integration 8GB Of HBM Outside , It also supports dual channels with a maximum of per channel 512GB Of DDR4 Memory .

Using the unique combination of hardware and software design , Without affecting the computing energy efficiency of the data center ,F1 DPU Provides maximum functional flexibility . This makes F1 DPU It can be used in environments with high performance density and low delay , Like storage (NVMe/TCP Storage uninstall )、 Security 、AI/ML(GPU decoupling ) And data analysis server (OLAP、OLTP Big data analysis engine ). Take storage , There is no need for x86 CPU and AFA In the storage system of ,F1 DPU It can be done 15M IOPS The performance of the , The bandwidth limitation here is entirely due to PCIe Its own bandwidth limit .

Official website ：https://www.fungible.com/

AWS

trace DPU The source of , Truly realize large-scale commercial DPU There are two major cloud computing giants ：Amazon AWS And Alibaba cloud .Amazon Nitro System from 2013 R & D started in ,2017 Official release , Designed to maximize performance and safety .

AWS Nitro The product family is designed to drive data center overhead （ Provide remote resources for virtual machines 、 Encryption and decryption 、 Fault tracking 、 Security policy and other service programs ） All from CPU Uninstall to Nitro On the accelerator card , Will release to the upper application 30% The original payment for “Tax” Calculation power .

Nitro The system is mainly composed of three parts ：

• With PCIe In the form of cards Nitro card , It mainly includes supporting network functions VPC（Virtual Private Cloud） card , Storage enabled EBS（Elastic Block Store）、Instance Storage Card and support system controlled Nitro Controller card .
• Nitro Security chip , The chip provides Hardware Root of Trust, Prevent software running on a generic server from pairing non-volatile storage Make changes , For example, virtual machine UEFI Program .
• Running on a universal server Nitro Hypervisor, This is based on kvm Lightweight of hypervisor, Mainly provide CPU And memory management , Simulation of equipment is not provided （ Because all devices are added to the virtual machine through transparent transmission ）.

Official website ：https://aws.amazon.com/cn/

Domestic manufacturers

Alibaba cloud

Alicloud on 2017 year 10 The DPCA architecture launched in May is regarded by the industry as the most successful DPU One of . Now , The fourth generation of Alibaba cloud dragon has begun to support Alibaba cloud's large-scale cloud business .

2022 At the Alibaba cloud summit in , Alibaba cloud has released a special processor for cloud data centers CIPU（Cloud infrastructure Processing Units）, Claim to replace CPU Become the cloud age IDC The processing core of .CIPU Relatively lightweight , Not a general-purpose computing chip , It is dedicated to the management and control of cloud computing data centers , It can be comprehensively dispatched CPU、GPU、 Storage hard disk 、 Switch and other hardware .

Ali cloud, CIPU And Amazon AWS Of Nitro Similar location . It is both a hardware box , It is also a control system , Docking with Feitian cloud operating system .CIPU It is mainly composed of special chip and controller , The form is like a box or smart card , It is mainly used to manage the flying cloud operating system .

• CIPU Downward access to physical computing 、 Storage 、 Network resources , Fast cloud and hardware acceleration ; Upward access to the flying cloud operating system , Manage and control millions of Alibaba cloud servers around the world ：
• CIPU Combined with calculation ： Fast access to servers of different types of resources , Bringing computing power “0” loss , As well as the hardware level security reinforcement isolation ;
• CIPU Combined with storage ： Hardware acceleration for block storage access of memory computing separation architecture , Cloud disk storage IOPS Up to 300 ten thousand , Long tail delay reduction 50%;
• CIPU Combine with the network ： Hardware acceleration for high bandwidth physical networks , Build large-scale resilience RDMA High performance networks , The minimum delay can reach 5us.

Official website ：https://www.aliyun.com/

Xin Qiyuan

Xinqiyuan was founded in 2015 year , Focus on network communication 、5G And cloud data centers , Customers include but are not limited to operators and secondary operators 、 Router switch equipment manufacturer 、OTT And Internet manufacturers 、 Network security vendors 、5G/6G Equipment suppliers, etc .

Xinqiyuan has completely independent intellectual property rights DPU chip . Xin Qiyuan DPU Compared with the traditional intelligent network card, it provides greater processing capacity 、 More flexibility 、 Programmable packet processing 、 Scalable Chiplet( Microchips ) Structure, etc . use NP-SoC Pattern for chip design , Universal ARM Architecture combines highly optimized packet oriented NP chip （RISC-V kernel ）、 Multithreaded processing mode , Make it possible to achieve ASIC Solidify the data processing capability of the chip , At the same time, the full programmable 、 Flexible and extensible properties , To support 400Gbps And above 、 Low power and cost-effective .

Xin Qiyuan DPU Adopted in the architecture Chiplet（ Microchips ） Technology is a new way of chip design , It is also the key chip technology that many enterprises in the industry are introducing .Chiplet Will meet the requirements of specific functions Die（ Naked piece ） adopt Die-To-Die Internal interconnection technology enables multiple module chips to be packaged with the underlying basic chip , Form a system chip .Chiplet Technology will be originally a piece of complex SoC The chip is decomposed into pellets , Similar to modular design , It helps to shorten the commercial time of products and the iteration of subsequent products , At the same time, it supports the connection with third-party chips Die-To-Die interconnection , It can also integrate more chips in specific professional fields . The performance and functional richness have been improved by leaps and bounds , It also provides core Qiyuan customers with more business scenarios .

Official website ：

https://www.corigine.com.cn/cn/index.html

Yisi core

Yisixin technology was founded in 2020 year 7 month , The team consists of domestic and foreign networks 、 In exchange for 、 Core storage professionals , On the Internet 、 In exchange for 、 Storage and high performance CPU And other fields have profound technical strength .

Stargate DPU Intelligent network card is the first commercial card in China 、 Having independent intellectual property rights P4 Programmable cloud native intelligent network card , Yisixin Technology P4 Network acceleration engine is the world's first vSwitch Designed for acceleration VLIW ISA P4 processor , It supports tens of millions of stream tables and can forward packets at wire speed . The network card is OVS、NFV、SDN vRouter、5G UPF The best choice for network application acceleration , High performance 、 Low latency 、 High flexibility 、 Low power consumption ：

• High performance ： On the network card , Single P4 The engine can realize the full duplex throughput rate of the network card .l
• Low latency ： The instruction level parallel processing architecture is adopted , The packet processing delay can be controlled at the nanosecond level .l
• High flexibility ： Full compatibility P4-16 edition , Meet the requirements of flexible protocol processing and smooth system upgrade .l
• low power consumption ： As a domain specific architecture （Domain-specific Architecture）, in the light of vSwitch Design for acceleration . Under the same performance index , Estimated power consumption is only traditional NP Architecture and multi-core CPU Architecturally 1/10.

Official website ：http://www.resnics.com/

Cloud vein core connection

Yunmai Xinlian was founded in 2021 year 5 month , It is a high-tech innovation enterprise focusing on cloud data center network chip product R & D and technological innovation .

2022 year 5 month 31 Japan , Yunmai Xinlian officially released the first domestic multi scene product independently developed RDMA Smart network card (DPU) product ——xFusion50.2023 In the first half of the year , CMAC will release the next generation of high performance DPU chip .xFusion50 It is the first product successfully independently developed by Yunmai Xinlian , It is also the first domestic implementation, including supporting end-to-end congestion control integrity RDMA Functional DPU product ,xFusion50 The programmable congestion control algorithm based on hardware implementation can effectively avoid network congestion , Give full play to RDMA Low latency and high performance of technology , Support cloud computing 、 High performance computing 、AI、 Full scenario deployment of storage cluster .xFusion50 The product has the following core highlights ：

• Support programmable congestion control algorithm , Programmable congestion control algorithm is the key technology to realize end-to-end lossless network ; It can also be through the open programmable underlying network interface , According to the networking characteristics of customers and the needs of upper layer services , Flexible support for a variety of congestion control algorithms , Maximize the traffic throughput of the business .
• Through independent research and development HyperDirect Technical support GPU Direct RDMA Is a cross compute node GPU Realize remote memory direct access , skip CPU To reduce the delay 、 Improve bandwidth , Improve the overall efficiency of distributed heterogeneous computing power clusters .
• Support network / Storage full scene unloading acceleration , Support vSwitch Full-offload , On the cloud VPC Network full function ; Support storage uninstall , Docking distributed storage NVMe-oF（TCP/RDMA）, Fully release the host CPU resources . And through support VirtIO Realize elastic network and elastic storage , Meet the business demands of seamless migration and rapid recovery of cloud users .

Official website ：https://www.yunsilicon.com/

Zhongkeyu number

Zhongke Yushu was founded in 2018 year , Focus on the R & D and design of dedicated data processors , Based on self-developed agile heterogeneous KPU Chip architecture and DPU Software development platform HADOS, The company independently developed the industry's first high-performance network and database integration acceleration function DPU Chips and standard accelerator cards , It can be widely used in ultra-low delay networks 、 Big data processing 、5G Edge of computing 、 High speed storage and other scenarios , Help computing become a new productivity in the digital age .

stay DPU Product R & D iterations , Zhongke Yu Yu Yu 2019 The first generation was released in DPU chip K1, Second generation DPU chip K2 Also in the 2022 The film was successfully put into production at the beginning of the year , At present, the third generation DPU chip K2 Pro R & D of ;2021 year 9 month , Zhongke Yushu starts the first round DPU Accelerator card products , At that time, the industry-leading 1.2 Microsecond . There is also DPU Storage accelerator card 、DPU Data computing accelerator card and other products and solutions are in the R & D process . In terms of the core technical features of the product , Zhongkeyu's DPU The chip innovatively uses software to define the accelerator technology route , Realize the software and hardware coordination DPU design scheme . The specific innovations are as follows ：

• Efficient heterogeneous multicore DPU framework , Define accelerator route based on software , Developed heterogeneous multi-core DPU Chip design method , Solve the problem of multi-core interconnection 、 Computational scheduling 、 Command control and other core issues .
• Ultra high bandwidth network protocol processing , R & D of private network protocol processing core and big data analysis and processing core , It solves the bottleneck of software parsing network packet protocol parsing and data processing , Greatly improve the communication efficiency between servers , Improve the horizontal scalability of the data center .
• Unified virtualization hardware platform , For data center networks 、 Calculation 、 Virtualization requirements for storage convergence , Research a unified and efficient hardware virtualization architecture , Solve the dilemma of single virtualization function of existing solutions （ Only network virtualization is supported ）, Release fully DPU Various resource capabilities , More efficient support for complex upper layer applications .
• A unified DPU Software development framework HADOS, Solve the problem of fragmentation of the existing programming framework , Make application deployment simpler and more efficient .

Official website ：

https://www.yusur.tech/zkls/zkys/index.html

Yu Zhixin

Dayu Zhixin was founded in 2020 year , Its founding and core team consists of domestic and foreign Internet companies 、 Cloud computing head companies and traditional networks 、 chip 、 It is composed of senior experts from safety head manufacturers , Have DPU Design and R & D and DPU Successful experience of large-scale commercial deployment .

Yu Zhixin Paratus series DPU The product adopts the parallel mode of three product lines to gradually launch easy-to-use and easy-to-use products for a wide range of commercial markets DPU product ：

• Paratus 1.0 As Dayu Zhixin DPU The first product line of , use ARM SoC As the main processing unit , Provide multiple 10Gbps/25Gbps Business network interface , At the same time, in order to facilitate user management , Separately set RJ45 Management .
• Paratus 2.0 As Dayu Zhixin DPU The second product line of , use ARM SoC + FPGA The hardware architecture of , stay Paratus 1.0 Product based , utilize FPGA High performance forwarding of data packets with solidified logic , Provide multiple 10G/25G、100G Business network interface .
• Paratus 3.0 As the third product line , Dayu smart core will be used for self-study DPU chip . The chip will be combined with the company's DPU Understanding of relevant technologies and future application scenarios , And the first two product lines （Paratus 1.0 and Paratus 2.0） Valuable customer feedback and experience gained in actual scenario deployment , Form a highly integrated DPU product .

Official website ：https://www.dayudpu.com/

Coming soon 《DPU The manufacturer's large stock market （ whole ）》！ Welcome to have technology 、 There are products 、 There are plans DPU The manufacturer shall actively declare SDNLAB Special planning “DPU The manufacturer's large stock market ” project .

Scan code to participate in “2022 year DPU The manufacturer's large stock market ”

Besides , The second SmartNIC&DPU The Organizing Committee of the Technological Innovation Summit has launched 2022 SmartNIC&DPU Awards Annual selection , welcome SmartNIC&DPU Excellent products and project practices in the technical field actively participate in , To promote SmartNIC、DPU Technological innovation and industrial development .

Issuance of detailed rules ！2022 SmartNIC & DPU Awards We sincerely invite you to participate in the annual selection

The second SmartNIC & DPU The Technological Innovation Summit was officially launched