Huawei's Arm server chip products made a stunning debut

On April 26, 2019, Huawei held the 2019 Huawei Intelligent Computing Conference in Hangzhou. With the theme of "Starting with the Chip to Make Intelligent Computing Pervasive", the conference released the heavyweight TaiShan solution, and gave a detailed and profound explanation of the TaiShan server from the three angles of "establishing the chip", "using the chip" and "releasing the chip". At the same time, it conducted in-depth trend insights and discussions on the AI ​​industry with authoritative experts from all walks of life.

Huawei focuses on investment in R&D. Among its total 180,000 employees, there are 80,000 R&D employees, who are widely distributed in more than 170 countries and regions. Its brand ranks 48th and it ranks 72nd among the Fortune Global 500.

With the advent of the Fourth Industrial Revolution, Huawei is committed to bringing the digital world to everyone, every family, and every organization, and building an intelligent world where everything is connected. Before 2018, Huawei's main strategy focused on connectivity, but after 2018, based on its understanding of the intelligent world, Huawei shifted to building a strategy of Connection + Computing + Cloud to build the key infrastructure of the intelligent world. Facing the intelligent world with hundreds of billions of connections and trillions of computing markets, Huawei has the hope and confidence to become one of the top five clouds in the world.

Wan Zhi, President of Huawei China Intelligent Computing Business Unit, pointed out that if connectivity solves the transformation from the physical world to the digital world, then computing will solve the transformation from the digital world to the intelligent world, and the cloud will make connectivity and computing more efficient.

Driven by Moore's Law, the industry is evolving from computing 1.0 to computing 3.0. Computing 1.0 was mainly dedicated computing for mainframes and minicomputers, computing 2.0 was general computing for data centers and X86, and computing 3.0 has entered the era of full-stack, full-scenario intelligent computing. The diversity of applications and data calls for an era of diverse computing, such as integer computing, floating-point computing, and other computing units. No single computing can cope with the diversity of big data. In the future, multiple computing architectures will develop together and flourish.

According to an IDC report, the penetration rate of AI enterprise applications was only 4% in 2018; however, GIV (Global Industry Outlook) predicts that by 2025, the cloudification rate will reach 85%, the data utilization rate will reach 80%, and the AI ​​application rate will reach 86%. The world is quietly moving from digitalization to intelligence.

However, today, the industry's intelligence is still facing four major challenges, which are constantly plaguing the development of this industry:

• Computing power is in short supply, scarce and expensive;

• Data cannot be coordinated, and central computing and edge computing cannot be connected;

• The deployment scenarios have high requirements and the environment is harsh and changeable, such as high altitude and extreme cold;

• The professional and technical requirements are high and there is a shortage of professional talent.

Huawei's layout in intelligent computing mainly includes four major areas: super computing power, cloud-edge collaboration, ubiquitous engineering and solutions.

(1) Super computing power: Huawei provides five self-developed chips to meet customer expectations. Super computing power is mainly reflected in the five aspects of computing, management, AI, storage, and transmission:

Computing: On January 7, 2019, Kunpeng 920, based on ARM processor chip, was released. It is the first 7nm ARM server processor that can provide 48/64 core multi-core and 2.6Ghz;

Management: The intelligent management chip Hi710 has a built-in intelligent management engine and intelligent fault management;

AI: In 2018, Huawei released the Ascend 310/910 AI chip based on its self-developed Da Vinci architecture, which has extreme energy efficiency and performance.

Storage: In terms of storage, it provides the self-developed intelligent SSD control chip Hi1812, which uses the intelligent acceleration and wear algorithm of PCIe NVMe and SAS to provide the industry's best chip in terms of performance and lifespan;

Transmission: The intelligent converged network chip Hi1822 supports Ethernet and PC convergence and is programmable, which can effectively reduce CPU resource usage and release at least 15% of CPU resources.

(2) Cloud-edge collaboration: Cloud models can be quickly pushed to the edge, and edge data can be quickly imported into the cloud. The process of data downlink and model backlink requires ultra-high bandwidth, ultra-low latency, and seamless coverage networks. Huawei can achieve data collaboration and interoperability through cloud-edge collaboration and seamless network coverage, allowing data to play a greater role.

(3) Ubiquitous Engineering: Huawei meets the requirements of various environment deployments through scientific research theories, GCTC test verification, and global deployment practices.

(4) Solution: Huawei's principle is to leave simplicity to customers and complexity to itself, and to provide integrated solutions for the industry. For example, in terms of data center integration, Huawei provides L1+L2 linkage/full liquid cooling cabinets to reduce the OPEX of data centers, and Atlas-based AI all-in-one machines can support 1024 AI clusters.

Huawei's intelligent computing is not just about servers. Huawei is implementing a strategy of developing three major computing platforms and two major ecosystems in parallel.

The three major computing platforms refer to Huawei's intelligent computing, which will be strategically deployed from the x86 computing platform, AI computing platform, and ARM computing platform:

X86 computing platform : By introducing two intelligent engines on the X86 service platform, namely intelligent management and intelligent acceleration engines, the FusionSeverPro server was released on April 3 this year. Its obvious features are flexibility, efficiency and quality.

Atlas: AI computing platform. Huawei hopes that this product can support the future of its entire AI. There are two major breakthroughs in Atlas. The first is computing power. Atlas is based on Huawei's self-developed Da Vinci architecture Ascend chip, which can provide super computing power. Among them, Ascend 310 is mainly used for edge computing and has the strongest edge computing power. Ascend 910 is mainly used for training scenarios and is currently the AI ​​chip with the highest single-chip computing density. The second is boundary breakthrough: Da Vinci architecture is self-developed by Huawei, so the biggest feature is the unified architecture, which is very easy to adapt.

The deployment of the ARM computing platform is the focus of this conference - the release of TaiShan server, which will be explained in detail below.

The "two ecosystems" refer to Huawei's continued efforts to build the AI ​​ecosystem and the ARM ecosystem : The AI ​​ecosystem will focus on business enablement, efficient cooperation, and developer and community building. The ARM ecosystem will expand its integrity from three aspects: application software, basic hardware, and basic software.

With the help of Huawei's powerful R&D capabilities and core chips, Huawei aims to achieve the automation upgrade of data centers, make AI affordable and effective for all industries, and drive the digital world towards the intelligent world. Currently, Huawei's products and solutions have been recognized by more than 10,000 customers.

The superiority of ARM architecture is self-evident. ARM architecture provides the best performance and power consumption ratio. Huawei's positioning of ARM is also very clear. ARM is an important part of Huawei's entire diversified computing. Huawei is also actively promoting the development of the ARM industry and the construction of the ecosystem.

Zhang Xiwei, President of TaiShan & Atlas, Huawei's Intelligent Computing Business Unit, summarized Huawei's ARM development strategy in three words: "establishing the core", "using the core", and "releasing the core".

Lixin: Huawei has been committed to creating new chips for many years and has independently developed the Kunpeng processor, which is the industry's most powerful ARM-based processor to date;

Using core: Huawei hopes to build a perfectly combined Taishan server based on the inherent advantages of ARM and Huawei's years of accumulation in the server field. This is Huawei's core work.

Fangxin: We are fully committed to the development of ARM's ecosystem and industry construction, accelerating innovation, and hope that partners can feel at ease working with Huawei.

Zhuangzi's "Xiaoyaoyou" says: "In the north sea there is a fish called Kun. Kun is so big that no one knows how many thousand miles it is. It changes into a bird called Peng. Peng's back is so big that no one knows how many thousand miles it is. When it flies in anger, its wings are like clouds drooping from the sky."

On January 7, 2019, the industry's highest-performance ARM-based Kunpeng 920 was released. The performance of Kunpeng 920 is as follows:

High performance, high integration, high throughput and high energy efficiency. Kunpeng 920 SPECint Benchmark score exceeds 930, which is 25% higher than the industry benchmark. It is worth mentioning that the chip integrates 8-channel DDR4, the rate has increased from 2663 to 2993, the total bandwidth has reached 1.5T bit/s, and the memory bandwidth exceeds the industry mainstream by 46%. Kunpeng 920 supports PCIe4.0 and CCIX interfaces, and can provide 640Gbps total bandwidth, which is 66% higher than the industry mainstream. The chip integrates 100G RoCE Ethernet card function, and the network bandwidth has quadrupled, greatly improving the system integration.

Xia Qin, deputy general manager of Huawei's processor chip business department, mentioned that in order to meet the diverse needs of customers, Kunpeng has launched a series of products. Among them, the Kunpeng 920 64-core can provide ultimate performance, the 32-core/48-core can provide ultimate energy efficiency, and the Kunpeng 916 32-core can meet the needs of low power consumption.

Kunpeng 920 reflects Huawei's deep accumulation in chips, and also reflects Huawei's strong appeal in the entire ecosystem. Kunpeng 920 has been designed with high performance as the basis, meeting customers' continuous demand for high performance and high energy efficiency. It adopts self-developed core architecture and self-developed chip architecture, and selects the industry's latest 7nm industrial ultra-large size packaging, fully enjoying the benefits of Moore's Law.

Kunpeng 920 was the first to launch the PCIe 4.0 standard, which greatly improved the throughput of various peripherals including storage and networking. At the same time, Huawei actively promoted the CCIX Alliance. The CCIX specification provides key technologies for heterogeneous acceleration, converting I/O access into memory access, greatly reducing access latency.

A journey of a thousand miles begins with a single step. Huawei has been exploring cooperation with ARM processors for 15 years. In 2004, it began to cooperate with ARM to develop communications and embedded devices; in 2009, it launched the smartphone chip K3, which is the predecessor of Kirin; in 2015, Huawei launched storage and server CPUs, and this series of chips is the predecessor of Kunpeng. In the future, Huawei will firmly invest in ARM and continue to evolve on Kunpeng processors, maintaining an evolution speed of at least one generation every two years to meet customer needs.

Huawei's intelligent computing has released five self-developed chips for computing, storage, transmission, management, and intelligence. TaiShan server is precisely the most attentive server product among Huawei's five self-developed chips for intelligent computing. The innovation of the five chips is perfectly reflected in the TaiShan server.

TaiShan Server has released three servers, namely the TaiShan 2280 balanced server. The so-called balanced server means that it is balanced in computing, storage and network, and is also the server with the most shipments and the widest application. The other is the TaiShan 5280 storage server, one cabinet can support 5.6PB of capacity, suitable for distributed storage scenarios; the third is the TaiShan X6000 high-density server, one cabinet can support 80 computing nodes and 10240 cores. The multi-core capability can support liquid cooling solutions, suitable for deployment scenarios of large data centers.

TaiShan server is equipped with two Kunpeng processors, which have strong computing power and become a new benchmark for server computing. In the SPECint benchmark test (mainly evaluating the integer computing power of the server), the TaiShan server's latest score is 1860 points, which improves the original record by 25%; in the SPECpower_ssj2008 test (mainly evaluating the balance between the performance and power consumption of the entire server system in the actual working environment), TaiShan server can improve the performance of Java applications in this test by 15% under full load.

Rome was not built in a day. TaiShan servers are the result of Huawei's 17 years of computing technology and process accumulation, multiple reliability designs and test guarantees, intelligent fault management in multiple scenarios, leading liquid cooling technology, etc., to ensure that the product is as stable as Mount Tai.

With the support of multiple technologies and multiple partners, TaiShan Server can provide five solutions for the entire site:

(1) TaiShan Big Data supports big data platform software, ARM multi-core can realize concurrent data processing, improve computing performance, support integrated deployment, and protect customer investment.

(2) Distributed storage: TaiShan supports multiple distributed storage software, such as SUSE, Ubuntu, Open-source, FusionStorage, etc. It can use the advantages of multiple cores to process data concurrently on a single server and improve storage performance.

(3) ARM native means that most mobile applications are developed based on ARM processors, which share the same instruction set as TaiShan servers and are naturally compatible. Based on this feature, Huawei has created cloud phones and cloud gaming scenarios. Currently, the performance of mainstream international cloud gaming service providers has increased by 30%. At the same time, Huawei ships 200 million mobile phones each year.

(4) HPC: In the past few years, Huawei has built a complete ARM HPC software station on TaiShan servers to help customers quickly deploy high-performance computing application environments, which have been deployed in key manufacturing, meteorological, and genetic industry applications.

(5) Database: TaiShan is currently pre-integrated with mainstream database software such as Nanda General, Renmin University of China Jincang, DAMO Database, MySQL, etc. Huawei has done a lot of software and hardware collaborative optimization to achieve the optimal computing path, with outstanding performance in improving query performance by 53% and transaction performance by 35%.

In terms of "chip" peers, Huawei will focus on industries, sectors, developers, and customers, make long-term investments, and build a three-dimensional ecosystem. The first measure is that Huawei will empower partners with a service attitude and method, and provide ARM development environment and tools with cloud services, including Huawei Cloud ARM verification environment, GCC Alliance Cloud Guizhou, Pengcheng Laboratory, etc. At the same time, Huawei R&D personnel will hold online and offline forums for joint incubation.

The second measure is to implement a partner incentive plan, including rights incentives for efficient scientific research organizations, industry ISV partners, and developer group rights incentives, to achieve win-win results with various companies.

"A single tree cannot make a forest, and only a thousand flowers can make spring." Lu Guang, general manager of TaiShan Server Business of Huawei Intelligent Computing Business Unit, said that only joint efforts can achieve the prosperity of the ARM industry.

Huawei's strategic layout in chips announces the arrival of an intelligent world where everything is connected and everything is aware. ARM-based smart terminal applications are accelerating. Huawei has accumulated a lot of experience in computing and continued to make innovative breakthroughs. Kunpeng spreads its wings and flies to Mount Tai~