Shulou(Shulou.com)06/02 Report--

If data is the means of production in the intelligent era and computing is the productivity in the intelligent era, then diversified computing chips are undoubtedly the power engine of productivity in the intelligent era, which determines the process of large-scale commercialization of artificial intelligence in the future.

At Huawei's full connectivity conference in 2019, Huawei announced its overall computing strategy for the first time. This means that at the turning point of the computing industry and the golden decade of computing architecture innovation, computing strategy will become one of Huawei's core strategies in the future. For example, Hou Jinlong, president of Huawei Cloud & AI products and services, said that by 2023, the global computing industry is expected to exceed US $2 trillion, and there is a market space of RMB 1.1 trillion in China. Huawei will unswervingly invest in the computing industry.

In fact, the computing market is indeed a huge market in the future. The emergence of a large number of artificial intelligence applications puts forward very high requirements for computing chips in terms of performance, power consumption, cost and application development, and also drives the development of computing towards diversification and heterogeneity. At the same time, the computing market still has extremely high technology barriers.

So, in the broad and competitive computing market, what is Huawei's core strategy for computing industry? What are the core competitiveness of Huawei computing industry? How will Huawei build an ecosystem in the face of computing fields that rely heavily on ecology for development?

From connection to calculation, why should Huawei arrange it like this?

In Huawei's view, two key technologies cannot be ignored in the intelligent era: one is connectivity and the other is computing. Calculation and connection are inseparable like twin brothers. Where there is connection, there should be calculation. Where there is calculation, there should be connection.

From the perspective of deeper computing demand, artificial intelligence, Internet of Things, big data and other technologies make application types diversified. Traditional general-purpose CPU has been "unable to do so" in the face of diversified application trend because it is subject to Moore's Law. General-purpose CPU computing is the bottleneck, but the advantage lies in scheduling, management and rich auxiliary tools; GPU computing is strong, suitable for high-concurrency application scenarios, but management is the bottleneck;FPGA management and computing are advantages, but the development cycle is long and difficult. Therefore, it has become a general trend to use diversified computing chips to flexibly meet the computing needs of diversified applications.

Hou Jinlong, president of Huawei Cloud & AI products and services, said that the computing industry is undergoing obvious trend changes. It is mainly reflected in: data centers are gradually evolving into computing centers, and the demand for AI computing power will increase by more than 10 times every year. By 2025, AI computing power will account for 80% of data center computing power.

Second, the terminal and edge end are driving the innovation of computing architecture. In 2018, the global shipment of Arm chips exceeded 23 billion pieces, mainly for the end side and edge side, while the shipment volume of data center general CPU was less than 30 million pieces. The huge demand and advantages of the end side and edge side will drive the innovation of computing architecture and the development of ecology, and the collaboration between the center side, end side and edge side in the future will bring more efficient computing.

In addition, the computing architecture must continue to innovate to solve the technical bottleneck problems encountered by the general CPU, and break through the problems of energy efficiency wall, heat dissipation wall, optimization wall, memory wall, high-speed IO wall and so on through diversified computing and heterogeneous computing, so as to make computing power more friendly to energy and environment. "Taking Huawei's newly launched Atlas 900 AI training cluster as an example, the computing power reaches 256 PFLOPS. Under the same computing power, the general CPU needs 6195 cabinets, the GPU needs 208 cabinets, and the NPU needs only 128 cabinets. In fact, we only use 16 cabinets through chip optimization and system-level optimization, and the power consumption is reduced from 40,000 kW to 736kW."

As the so-called trend of computing diversification is unstoppable, Huawei has begun to unswervingly invest in the computing industry. In fact, Huawei has a long-term accumulation in the field of computing chips. Since investing in the research and development of the first embedded chip in 2004, Huawei has become the only manufacturer in the industry with five key chips of "CPU, NPU, storage control, network interconnection and intelligent management" in computing architecture.

At this Huawei Full Connection Conference, Huawei announced that it would build an industrial layout of "one cloud, two wings, two engines + open ecology" around the dual engines of "Kunpeng + Shengteng"; The two wings refer to intelligent computing business and intelligent data and storage business, covering end, edge and cloud scenarios, providing diversified computing power of "Kunpeng + Shengteng +X86+GPU," optimizing the cost per bit of data around the whole life cycle of data and maximizing the value per bit of data; One cloud refers to Huawei Cloud, which becomes the black land of ecological partners through full-stack innovation, providing inclusive computing power for the world; finally, it opens hardware and software to enable partners to form an open ecosystem.

To this end, Huawei will fully promote Huawei's computing industry strategy from four aspects: architectural innovation breakthrough, investment in full-scene processor family, adherence to the business strategy of doing something and not doing something, and spare no effort to build an open ecosystem. "Huawei's computing strategy pursues continuous innovation around customer needs, providing the world with the strongest computing power, making clouds everywhere, and making intelligence omnipresent. "Hou Jinlong bluntly said. It is reported that Huawei will invest in chips according to the rhythm of "mass production generation, R & D generation and planning generation," and will continue to evolve and be backward compatible.

"Kunpeng + Shengsheng," what kind of double lift Huawei wants to build

There is no doubt that in the computing field, Huawei will build its core competitiveness on Kunpeng processor and Shengteng processor, build an ecological whole chain around Kunpeng processor, and gradually verify the system level from motherboard, server and server operating system to cloud service and application; Around Shengteng processor, full-stack full-scene AI solution has been built. Shengteng 310 last year and Shengteng 910 this year have been announced for commercial use for six consecutive times, marking the landing of full-stack full-scene AI solution.

It is reported that Huawei has invested more than 20,000 engineers in chip-related research and development. In fact, Kunpeng processor has been developed since 2007. After 12 years of research and development, it is now the third generation chip product. Huawei revealed that Kunpeng + Shengteng will evolve for a long time, continuously improving competitiveness at the pace of launching a generation every year and ensuring backward compatibility.

In terms of Kunpeng processor, Huawei launched TaiShan series, a universal server based on Kunpeng processor and with the strongest computing power. TaiShan series servers have at least 64 cores, 8 memory channels, PCIe 4.0, all-in-one SoC, xPU high-speed interconnection, 100GE high-speed I/O and other six features, which are currently the most powerful general purpose servers compatible with Arm architecture. At the same time, Huawei provides storage-intensive, compute-intensive, edge computing and other server products for the needs of distributed evolution of data centers and edge computing.

In terms of Shengteng AI processor, Huawei released Atlas full series products with the strongest computing power in the world based on Shengteng AI processor, including Atlas 900, AI training server Atlas 800, intelligent station Atlas 500, AI reasoning and training card Atlas 300 and AI acceleration module Atlas 200, which are the fastest AI training clusters in the world based on Shengteng 910 and 310 AI processors. The Atlas full series product layout has been completed, covering cloud, edge and end scenarios. Powerful computing power for training and reasoning. Atlas series products are based on unified Da Vinci architecture and full-scene AI computing framework, which can realize cloud edge collaboration, which is conducive to intelligent acceleration of the whole industry.

Atlas 900 is composed of thousands of Shengteng 910 processors and ranks first in the world with 59.8 seconds in ResNet-50 test. It can be widely used in scientific research, astronomical exploration, oil exploration and other fields. Atlas 800 integrates 8 Shengteng 910 AI processors in 4U space, which can provide super computing power of 2 PFLOPS, and the computing power density is 2.5 times that of similar products in the industry. Atlas 300: It can provide 256 TFLOPS of computing power, twice the current mainstream training card in the industry, and the number of training pictures per second has increased from 965 to 1802. "Ma Haixu, president of Huawei's intelligent computing business unit, said.

In fact, according to the information disclosed by Huawei at this fully connected conference, Huawei has initially completed the layout of Kunpeng and Shengteng at the product and service level. At the product level, Huawei has a wealth of products for users to choose from, whether it is general server products or AI training clusters; In terms of cloud services, Huawei has also launched 112 cloud services based on Kunpeng and Shengteng, including 69 cloud services based on Kunpeng and 43 cloud services based on Shengteng, which can enable more users to enjoy the computing power brought by Kunpeng + Shengteng through cloud services. Hou Jinlong revealed: "Huawei Cloud Kunpeng Cloud Service has the characteristics of high performance and high concurrency; Shengteng Cloud Service can be widely used in AI reasoning, training, automatic driving training and other scenarios to realize inclusive AI through full-stack innovation. "

Of course, in addition to the product and service level, Huawei will also launch a series of measures from software and hardware, ecological cooperation and other aspects to promote the development of Huawei computing.

From hardware to software to partners, how does Huawei promote ecology?

As we all know, the computing industry is an industry that is extremely dependent on ecology. In addition to the chip products that need to be competitive enough, it must have strong ecological capabilities in the whole industry chain, so that the related computing industry can get considerable development. From the development history of chips for more than 60 years, an open computing ecosystem can finally prosper.

To this end, Huawei launched two major initiatives, hardware open and software open source, at this fully connected conference to enable partners.

First, at the hardware level. Huawei officially launched the Kunpeng motherboard, which is fully open to partners. Huawei Kunpeng motherboard uses xPU high-speed interconnection, all-in-one SoC, 100GE high-speed I/O and other key technologies. It is not only equipped with Kunpeng processor, but also built-in BMC chip, BIOS software. Huawei will open up motherboard interface specifications and equipment management specifications, provide reference design guidelines for complete machines, and fully open Huawei's technology accumulation and practical experience to partners. Partners can quickly develop their own branded server and desktop products based on Kun Peng motherboards and complete reference design guides.

Second, at the software open source level. Huawei announced that it will open source the server operating system on December 31, 2019, and open source GaussDB OLTP stand-alone database in June 2020. The open-source operating system name is openEuler, and the open-source database name is openGauss. OpenGauss database can cover more than 70% of database business scenarios of enterprises. Huawei will support openEuler-based partners to distribute commercial versions of operating systems, and mainstream applications and software developers in various industries to migrate software and applications to openEuler-based operating systems. It is reported that Huawei and Deep Technology, winning the bid Kirin, Tianjin Kirin, China Mobile Suzhou Research Institute, Puhua and other partners plan to jointly launch openEuler open source community.

Finally, at the enabling partner level. Huawei announced that it will invest US $1.5 billion in the development of industrial ecology in the next five years to better support ecological development, including aggregating applications by industry and building complete computing industry ecological chain and competitive solutions around various industries; integrating industries by region and building its own Kunpeng industry according to the characteristics of regional partners; Incubate standards with alliances, build standard systems for basic software and hardware through green computing industry alliances, edge computing industry alliances, etc., and promote the healthy development of the industry; develop developers with communities. Huawei hopes to train 5 million developers with partners in the next five years.

In summary, Huawei hopes to achieve the products and services in place, the industry support funds in place, the industry standards in place, the talent training in place through a series of measures, and build a prosperous Kunpeng + Shengteng computing ecosystem with partners.

The future journey is the sea of stars

Computing industry is a typical technology-intensive and capital-intensive industry. In the early stage, huge investment in R & D is often required, and innovation ability needs to be continuously maintained to ensure competitiveness. At present, after many years of stability, the global computing industry pattern has ushered in a new round of changes. Computing architecture innovation has begun to enter a new golden age. The computing market will be a market with huge space and unknown competition.

Huawei's first 30 years focused on things between connections and achieved great success. Today, in the era of intelligence, connectivity and computing are inseparable. Huawei chooses to start again with computing industry. Challenges and opportunities coexist. Hope Huawei can live up to Shaohua, dream for horse, star sea.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.