Today, the return of the offline Yunqi Conference will undoubtedly detonate the entire science and technology circle. If any company can make its own activities into the “evening party” of the entire science and technology circle, it must be the Alibaba Yunqi Conference.
At the main forum this morning, Zhang Jianfeng, Dean of Alibaba Dharma Academy and President of Alibaba Cloud Intelligent Business Group, gave a speech on the theme of “Deep Clouds, New World”. If there is a company in the world who believes in “cloud”, it must be Alibaba. Zhang Jianfeng believes that “cloud was originally a part of IT, but now IT has become a part of cloud.”
Alibaba has always adhered to the “one cloud, multiple cores” strategy, shielding the differences in hardware downwards and providing consistent services upwards. And in this most important link, Alibaba’s semi-conductor company Pingtou also released the industry’s long-awaited self-developed cloud chip-Etian 710.
The best ARM server chip in the industry
According to Zhang Jianfeng, Yitian 710 is based on Arm’s latest ARMv9 architecture design and adopts the industry’s most advanced 5nm process. The number of transistors in a single chip is up to 60 billion. It can be called the “most capable” Arm server chip in the server chip industry.
Because the 5nm process puts forward extremely high requirements on the energy density and the layout of the internal structure of the chip. For this reason, Pingtou flexibly dispatched 30 different EDA software, deep customized clock network and customized IP technology during the research and development process. In addition, they also adopted advanced multi-chip stacking technology, and finally successfully ensured the optimization of chip performance and power consumption.
In response to the requirements of high concurrency, high performance, and high energy efficiency in cloud computing scenarios, Pingtou has done in-depth customization work for Etian 710, and also introduced many self-developed new technologies to integrate leading chip design technologies with cloud scenarios. The unique requirements are combined to achieve a breakthrough in performance and energy efficiency ratio.
In order to solve the bandwidth bottleneck under the condition of a large number of cores, Pingtou has made special optimizations on the on-chip interconnection, using new flow control algorithms to reduce system backpressure, effectively improving system efficiency and scalability, and effectively transforming single-core high-performance into The high performance of the entire system. In addition, through the new conversion mechanism from system address to DRAM address, Etian 710 supports security, non-secure isolation, multiple NUMA, and abnormal channel isolation features, greatly improving DRAM read and write efficiency.
Etian 710 contains 128 CPU cores, with a clock speed of up to 3.2GHz, while taking into account performance and power consumption. In terms of memory and interfaces, Etian 710 integrates the industry’s most advanced DDR5, PCle5.0 and other technologies, which can effectively improve the chip The transmission rate is suitable for various cloud scenarios.
Zhang Jianfeng introduced that on the world’s authoritative CPU benchmark test set SPECint2017, Etian 710 scored 440 points, the performance exceeded the industry benchmark by 20%, and the energy efficiency ratio increased by more than 50%, which can effectively help data center energy conservation and emission reduction.
Alibaba Cloud manages more than 1.5 million servers worldwide. These large-scale clustered servers have also brought many problems. Alibaba has continuously reduced overall computing costs and solved high energy consumption issues through self-developed technologies and innovations. For example, in Zhang The data center in North is powered by natural wind all year round, and the data center in Hangzhou uses liquid cooling to solve the problem of high energy consumption.
In addition to advancing rapidly at the server chip level, Alibaba also released the Panjiu server on the spot.
Panjiu’s self-developed server series is oriented towards the cloud-native era. It is the first server series equipped with the self-developed chip Etian 710, which combines high-performance computing and high-performance storage. This server will be deployed this year for Alibaba Cloud’s own use.
Zhang Jianfeng introduced that the Panjiu server series adopts a flexible modular design, which can realize the separation of computing and storage, including high-performance computing series, large-capacity storage series, and high-performance storage.
With the release of “Yitian” and the launch of “Panjiu”, Alibaba Cloud has also perfected the last link of the full-stack cloud infrastructure, achieving technological and architectural innovation and self-research from chips, components to complete machines.
Cloud computing provides opportunities for overtaking in corners
Etian 710 is an important step for Alibaba Cloud to promote the “One Cloud, Multiple Cores” strategy, and it also reflects Alibaba’s determination to overtake in corners.
Alibaba has always maintained a high-profile attitude in core manufacturing, and has started layout very early.
In terms of investment layout, in 2016, Alibaba invested in software-defined network (SDN) chip companies Barefoot, Aojie Technology, Cambrian, Shenjian, Nene and other chip companies. In October of this year, Alibaba and Baidu also strategically invested in Feiteng Information Technology Co., Ltd., a domestic CPU developer based on the Arm architecture.
In terms of self-research, Alibaba established Dharma Academy in 2017 and formed a technical team composed of top experts in the semiconductor industry. In 2018, Alibaba acquired wholly-owned Zhongtian Micro, the only independent embedded CPU IP core in mainland China, and integrated the self-developed chip business of Dharma Institute at the Yunqi Conference in the same year to become the strongest chip company in Alibaba. elder brother”.
In 2019, Pingtou launched the first AI inference chip “Hanguang 800” and entered mass production. Starting in 2020, it has been deployed on a large scale in Ali’s super data center.
Beginning in 2009, after more than ten years of development, Ali has established a large and complete software and hardware ecosystem. And this time Ali entered the field of self-developed server CPU, and behind the release of Etian 710 may be the restlessness and offensive launch of the entire Arm ecosystem.
From the perspective of servers, the lowest level equipment of the cloud, X86-based servers have long dominated the market for a long time, and a mature business ecosystem has been built, occupying the absolute right to speak for patents and standards. According to data from Tianfeng Securities, in the server market, the domestic X86 architecture market accounts for 96.4% of the market, which is basically monopolized by Intel. Since the server turning point in the second half of 2018, the Arm architecture quickly occupied 0.9% of the domestic market. Share.
The rapid development of the Arm architecture has also brought cloud vendors more enthusiasm to a large extent. In the past two years, manufacturers including Ampere, Fujitsu, Mavell, Amazon, and Huawei have all made efforts to make cores. In November 2018, Amazon AWS, the world’s largest cloud service provider, launched the first AWS Graviton server chip based on the Arm architecture, and in December 2019 launched the second-generation 7nm Graviton server chip with Arm Neoverse N1 core.
In China, Huawei’s HiSilicon has launched the 7nm 64-core server chip Kunpeng 920 based on the Arm architecture as early as January 2019, as well as the server “Taishan”. However, subject to sanctions, the next-generation “Kunpeng 930” could not be released as scheduled; Baidu released the AI chip “Kunlun” in 2018, and it is reported that Baidu’s second-generation Kunlun chip has been successfully taped out and will be mass-produced in the second half of 2021. .
Although Ali Yitian 710’s “out of the sheath” is late, and it is the industry’s first 5nm chip, Ali still sees unlimited hope.
However, Zhang Jianfeng said that Yitian 710 will not be sold, mainly for Alibaba Cloud’s own use, which is an important step in Alibaba Cloud’s “one cloud with multiple cores” strategy. “We will continue to maintain close cooperation with partners such as Intel, Nvidia, AMD, Arm, etc., to provide customers with more choices.”
Ali’s full stack layout and ambition
In three years, Pingtou completed the “triple jump” from the release of the first RISC-V processor Xuantie 710, the first cloud AI inference chip Hanguang 800 to the first general-purpose server chip Etian 710.
In addition to today’s protagonist Yitian 710, at the Yunqi Conference, Zhang Jianfeng announced that Xuantie CPU has shipped more than 2.5 billion units, becoming the largest domestically-produced CPU in China. The Xuantie series of processors are self-developed CPUs developed by Pingtou for IoT end-side applications. It adopts two architectures, self-developed and RISC-V, covering various scenarios from low power consumption to high performance. Xuantie CPUs are widely used Machine vision, industrial control, vehicle terminal, mobile communication, multimedia and wireless access fields.
At this conference, Alibaba Cloud also announced the open source of four Xuantie RISC-V series processors, Xuantie E902, E906, C906 and C910, covering high, medium and low application scenarios, and opening related tools and system software. . Developers around the world can download the source code of Xuan Tie through Github and the Open Chinp Community (Open Chinp Community).
In addition to Ali’s first large-scale application, the cloud AI reasoning chip “Hanguang 800” and today’s “Yitian 710” realized by customers in industries such as search recommendation and video live broadcast through Alibaba Cloud services, Ali Pingtou already has the processing For product families such as device IP, AI chips, and general-purpose chips, its end-cloud integration strategy is becoming clearer.
With the development of 5G communication technology, the migration of computing and data to the cloud will accelerate in the future, which will give birth to more new species on the cloud. Zhang Jianfeng believes that just as electricity was first available and then the power grid was established, after the completion of such infrastructure construction, a wealth of electrical appliances appeared, which changed our life and production methods.
The cloud is the same. There will be countless new species on the cloud in the future. Zhang Jianfeng said, “Such a technological explosion has a solid foundation. We have completed the construction of infrastructure and will soon see the same era opportunity as the explosion of electrical appliances. “