Electronic enthusiast network reported (Wen / Huang Jingjing) since April this year, arm disclosed the technical details of neoverse V1 and N2 platforms, the new products and cooperation progress of partners in the market have been continuously promoted, covering cloud services, high-performance computing, 5g / network, etc. Recently, Zou ting and frank Zou, senior global director of arm Infrastructure Division, introduced the specific cooperation progress to electronic enthusiasts and other media, and also shared arm’s strategy and planning in the field of server chips.

CPU based on arm neorse platform has significant performance and cost advantages

Oracle released ampere A1 compute and introduced Oracle cloud infrastructure (OCI) based on the example of arm architecture. Among them, ampere Altra launched the industry’s first 80 core arm architecture server SOC and the first cloud native CPU for modern cloud and edge computing data centers.

 

Ampere Altra is the “first cloud native processor” released by ampere computing in March this year. It has 80 64 bit arm processor cores. The next generation ampere Altra Max will have 128 cores, be compatible with ampere 80 core Altra processor and support two-way server platform.





Oracle’s new OCI ampere A1 compute cloud instance has a cost advantage of US $0.01 per core per hour and a highly flexible configuration that can expand 1 to 80 CPUs. It provides additional options for developers when developing high-performance computing applications based on next-generation arm architecture such as video coding, highly extended network applications such as mysql, or machine learning inference. Zou Ting said that such cost advantages and scalability are very attractive to customers.

 

Zou Ting also introduced that cloudflare is a service provider providing CDN content distribution network and mitigating DDoS Attacks Based on network infrastructure. In July this year, the company adopted arm architecture server for deployment for the first time. By adopting CPU based on arm Neover platform, the performance per watt can be improved by 57%. Compared with the server designed in 2013, cloudflare can safely handle more than 10 times of internet requests on the system based on ARM architecture.

 



 

In China, Ali cloud, Baidu Intelligent Cloud and Tencent are developing database based on Arm platform.

 

For example, Alibaba cloud has launched the public beta of ECS arm instance specification family, and the generic instance g6r based on the arm architecture processor and the third-generation DPCA architecture has been launched; Baidu Intelligent Cloud database is actively developing Arm platform database products, has completed the Arm stack full code transformation and performance optimization, performance of traditional architecture platform. Through performance test, data exchange, technical discussion and other forms, Tencent and arm continue to promote the development of JDK in arm architecture. At present, ZGC on Tencent has been mass produced based on ARM architecture and has been applied in practice.

 



 

Fuyue supercomputer based on ARM architecture ranked first for three consecutive times

The supercomputer “Fuyue”, jointly developed by RIKEN and Fujitsu based on the arm architecture, ranked first in the international supercomputer top 500 list for the third time in a row. This achievement highlights that the arm architecture can meet the complex computing needs in the field of high performance computing (HPC) with unparalleled energy efficiency, performance and scalability. As more solutions based on ARM architecture are adopted, the arm HPC ecosystem is growing rapidly. Many partners have announced plans to use the recently launched neoverse V1 platform for megabyte and HPC projects, including sipearl, Korea Institute of electronics and communications (ETRI) and meity, Ministry of electronics and information technology of India.

5g / network: release of the first network solution using neoverse N2

Marvell released octeon 10 DPU based on neovse N2 platform. Compared with previous generations of octeon products, octeon 10 performance is improved by 3 times and power consumption is reduced by 50%. This newly released solution is not only the first network solution using neoverse N2, but also the first kernel using 5 nm process.

 

NVIDIA announced that it will expand its support for ARM architecture CPU on the NVIDIA aerial A100 ai-on-5g platform. The computing platform uses the NVIDIA aerial software development kit and will work with NVIDIA bluefield-3 A100 with 16 built-in arm cortex-a78 processors to provide high-performance 5g ran and AI platforms for enterprises and network equipment manufacturers.

 

In China, through engineering cooperation, resource docking, ecological promotion and other forms, arm continues to help Chinese operators build a new generation of 5g + aicde intelligent infrastructure.

 

The development of these application fields is not consistent. Zou Ting said that in the cloud field, arm has achieved relatively leading success. AWS, Oracle, Tencent and Baidu are all designing and optimizing their application software, operating system and middleware based on ARM architecture.





In terms of HPC, Fuyue’s supercomputer has once again achieved great success in HPC performance testing.

 

In terms of 5g, domestic Lenovo, rihai, and foreign Nokia and Samsung have adopted ARM technology. Although edge computing is a relatively mature market, it also has partners including NXP, Marvell, NVIDIA and so on. It is expected that arm will also achieve considerable success in this field in the future.

 

The arm server chip is unstoppable

A few years ago, the expansion of arm in the server market may not be smooth. However, from the more and more adoption of the above arm neoverse platform by large Internet manufacturers, we can see that arm server chips have received unprecedented attention. Why?

 

Zou Ting believes that the traditional universal solution is no longer applicable to the diversified requirements of infrastructure in the future. Facing more and more diversified workloads, the slowdown of Moore’s law and meeting the requirements of data center floor area and power density, the innovation being promoted by cloud service providers and super large Internet companies has also reached an unprecedented level. These are driving the optimization of the entire data center. Now these optimizations are gradually migrated from the upper application to the whole machine and extended to the server processor.

 

Through neoverse, arm and a wide range of ecological partner systems are providing a unique platform to realize future solutions for server chips. Among them, the strong development momentum brought by AWS graviton2 is a very powerful proof of this common vision.

 

In the Chinese market, domestic neoverse customers and partners are discussing and cooperating with us in the field of server application. I believe that in the near future, domestic partners will release more and more   Dynamics and progress of products and solutions of arm neoverse platform.

 

Internet manufacturers choose self-developed chips out of the demand for customized chips, while for those independent arm server chip manufacturers, they prefer standardized arm server chips. Zou Ting said that a large number of enterprise customers are very interested in the high cost performance and excellent TCO performance brought by the standardized arm server chip. In particular, the rapid growth in the demand for standardized arm server in the second-line Internet and enterprise market has brought very broad market prospects to independent arm server chip manufacturers.

Arm neorse is the preferred option for DPU

Earlier, Marvell released octeon 10 DPU based on neovse N2 platform. NVIDIA announced that ARM technology is used in bluefield-3 DPU, and Mount Evans released by Intel adopts 16 arm neorse N1 cores. In China, the development of DPU has also become a hot field.

 

According to Zou Ting’s analysis, from the perspective of the implementation of DPU, DPU is connected to the host system and receives power through PCI Express slot, which will limit the available energy consumption of DPU. For this reason, in addition, DPU supports special workload accelerator and packet processing functions, it is very important whether the processor in DPU can provide excellent performance per watt. and   Neoverse n series can provide industry-leading performance per watt, which is used as DPU   The best choice for the core.

 

He also said that not only the domestic Internet giants, but also the world’s leading Internet companies, while gradually deploying and adapting standardized DPU products, most are also exploring self-research or outsourcing to develop customized DPU products suitable for their special business needs. ASIC SOC DPU is the universal final target product form, Although in the middle, we also see the solution of transition with FPGA.

Different strategies of the three product families of neoverse platform

According to Zou Ting, arm neoverse is divided into three product series, and each product series has its unique advantages for different application fields. The V-series platform emphasizes a new computing series with performance as the priority. At present, neoverse V1 platform is the representative.





N series platforms focus on PPA and find a balance design among performance, power consumption and area. At present, the products of this series include neoverse N1 and N2 platforms. Suitable for CPU chip of general server, application chip of intelligent network card, network equipment of 5g base station, etc.

 

E-Series platform: the main focus is throughput efficiency. It is optimized in terms of power consumption and area reduction. It is very effective for network traffic and data applications. At present, neoverse E1 platform is the representative.

 

In addition, arm neoverse cmn-700 is the most advanced mesh interconnection technology in the industry, which can give full play to the performance and per watt performance advantages of neoverse V1 and N2 platforms. In this regard, Zou Ting said that the latest neoverse kernel supports new mathematical functions and data formats, which can be used to improve the reasoning performance of machine learning. Compared with traditional CPU, these functions can accelerate reasoning throughput up to 4 times. For machine learning training, GPU and NPU are generally used to perform most processing tasks. However, CXL can be supported in arm cmn-700 interconnection technology to realize the connection between these GPU and NPU and CPU, and provide connection with lower delay and higher throughput.

China’s infrastructure market has great opportunities

Zou Ting said that in the Chinese market, we focus on the two vertical fields of cloud data center / super large Internet and 5g network / telecom market. As mentioned above, there are already many domestic partners.





In the wave of localization of semiconductor chips, there are hundreds of chip industry tracks, such as CPU, DPU, GPU, etc. a large number of startups with development potential have emerged in these fields. Arm neoverse IP platform provides a very competitive cornerstone for chip development in the field of infrastructure. At the same time, arm is willing to work with domestic chip manufacturers to jointly help domestic industrial upgrading through technical exchange and ecological docking.

 

 

Leave a Reply

Your email address will not be published. Required fields are marked *