Shen Yao's Technology Observation
Understand technology and win the future!
A few days ago, at the Fifth World Internet Conference held in Wuzhen, Huawei's Ascend 310 artificial intelligence chip stood out from more than 400 declared scientific and technological achievements and won the Fifth World Internet Leading Scientific and Technological Achievement Award. It can be said that among the 15 leading scientific and technological achievement awards selected by this conference, the importance of Huawei Ascend 310 is enough to rank among the top three, and it deserves this honor.
I still remember the shocking scene at the Huawei Connect Conference held in October this year, when Xu Zhijun, the rotating chairman of Huawei, directly showed the Ascend 310 chip and responded to the previous rumors about Huawei's AI chip: "The outside world has been rumored that Huawei is developing AI chips. I want to tell you that this is true."
As the saying goes, if you don't make a song, you will be a blockbuster. Huawei closed its doors to "do things" and shocked the world as soon as it made a move. Today, with the Ascend chip as the "base", Huawei has released a full-stack, all-scenario AI solution. The full stack aims to provide AI application developers with powerful and economical computing power and a low-threshold application development platform, making AI data modeling, model training, and application development simpler, more agile, and more efficient. Full-scenario means that intelligence is omnipresent, covering any business scenario of the end, edge, and cloud.
At the same time, as an important part of the full-stack, all-scenario AI solution, Huawei also released the Atlas intelligent computing platform. With the support of the Ascend chip, the Atlas intelligent computing platform delivers AI capabilities to the device side, edge side, and data center side, making computing ubiquitous, and thus realizing ubiquitous AI.
AI must start from the "core"
Economists believe that human development has produced a total of 26 general purpose technologies (General Purpose technology), and artificial intelligence is one of them. Simply put, a "universal technology" is one that has multiple uses, can be applied almost anywhere in the economy, and has large spillover effects.
Although AI is still in the stage of weak AI or narrow AI for specific applications, it will eventually evolve to strong AI and pan AI. According to Huawei's prediction, by 2025, there will be 40 billion smart terminals worldwide, the penetration rate of smart assistants will reach 90%, and the data usage rate of enterprises will reach 86%. Intelligence will be as ubiquitous as the air.
In other words, AI, as a new general-purpose technology, will transform every industry and every organization. This requires the entire industry to make changes in ten aspects, including future model training, computing power, security, algorithms, automation, applications, technology collaboration, platforms, and talent acquisition. These ten changes are an important foundation for the future development of AI, and they are also Huawei's expectations for the development of the AI industry and the driving force behind its own AI strategy.
AI has brought new opportunities for China's industrial upgrading. Thanks to the support of Chinese policies, the introduction of a large amount of capital, a wide range of user needs and massive data, AI has entered the fast lane in China. According to the "White Paper on China's Artificial Intelligence Industry" released by Deloitte, the scale of China's artificial intelligence market is expected to exceed 30 billion yuan in 2018; the number of artificial intelligence enterprises in China exceeds 1,000, ranking second in the world.
Nevertheless, China's artificial intelligence market is still in its early stages of development. Although it is second only to the United States in terms of investment, papers, and number of companies, it still lags far behind the United States in basic research fields such as chips and algorithms, as well as key indicators such as talents and ecology. This is bound to become the biggest stumbling block in the development of China's industrial AI.
The advent of Huawei's Shengteng chip has undoubtedly injected a boost to the entire AI market. Of course, the more important significance is that with the support of chips, AI can be implemented in a wider range of scenarios, from terminals, edges, to data centers, and truly take a big step towards "universal technology".
Make AI truly a general-purpose technology
Huawei released two chips, the Ascend 910 and 310, at this year's Fully Connected Conference, based on Huawei's self-developed "Da Vinci" computing architecture. Among them, the Ascend 910 has the highest single-chip computing density, which is twice as high as the 125T of the most powerful NVIDIA V100 at present, and is expected to be officially launched in the second quarter of next year; the Ascend 310 has already been mass-produced, and it is an AI SoC with extremely efficient computing and low energy consumption, focusing on terminal low-power AI scenarios.
Huawei Ascend 310 is the first full-stack, all-scenario artificial intelligence chip. Under the typical configuration of 8W power consumption, it can output 16TOPS@INT8 or 8TOPS@FP16 performance, supporting the data accuracy requirements of training and reasoning scenarios. Built-in a variety of hardware accelerators, such as DVPP module, compression and decompression engine, encryption and decryption engine, etc., taking into account the calculation acceleration and performance upgrade. It also provides rich interfaces and a one-stop AI integrated development kit to reduce the difficulty of AI development.
According to the test, Huawei Ascend 310 can support simultaneous recognition of 200 different objects including people, cars, obstacles, and traffic signs, and can process thousands of pictures in one second. This capability allows the Ascend 310 to show its talents in various scenarios such as smart cities, autonomous driving, smart new retail, robots, industrial manufacturing, and cloud computing AI services, providing high-efficiency computing power within reach for all walks of life.
In October this year, Huawei and Audi demonstrated the innovative achievements of L4 driverless driving. Audi cars are equipped with Huawei's MDC mobile data center. According to calculations, the Ascend 310 chipset only consumes 200 watts of energy when supporting L4 unmanned driving, which is a very complex edge computing scenario.
Looking back on the development of artificial intelligence in the past sixty years, insufficient computing power is a key factor that prevents the large-scale and commercial application of AI. Although the world's fastest computer has achieved a computing power of 2 billion billion operations per second, general-purpose chips still cannot meet the requirements of low latency, low energy consumption, and high computing power in many scenarios on the device side and edge side, which restricts the implementation of AI in a wider range of scenarios.
The advent of the Ascend chip can greatly accelerate the practical application of AI in various industries, and also marks that Huawei's AI solution has achieved industry leadership at the underlying chip level. In addition to chips, Huawei's full-stack, all-scenario AI solution also includes a chip operator library and CANN, a highly automated operator development tool; MindSpore, a unified training and inference framework that supports end, edge, and cloud independence and collaboration; and ModelArts, which provides a full-process service. Huawei not only provides the computing power required for all-scenario AI, but also provides methods to apply computing power to all scenarios.
Huawei Atlas makes computing ubiquitous and intelligence ubiquitous
At the 2018 Huawei Connect Conference, Xu Zhijun, Huawei's rotating chairman, said that Huawei's AI chips will not be sold separately, but will be delivered in the form of acceleration modules and server integration modules. It is not difficult to understand that the advantage of this delivery model is that Huawei can provide customers with the most optimized solution to maximize the performance of AI chips, and at the same time lower the threshold for customers to develop and apply AI, which helps to implement products in a wider range of scenarios.
Therefore, based on the Huawei Ascend 310 chip and the industry's mainstream heterogeneous computing components, Huawei launched the Atlas intelligent computing platform at the 2018 Fully Connected Conference.
Specifically, the Huawei Atlas intelligent computing platform mainly includes several major product series:
First, the Atlas 200 AI acceleration module: only half the size of a credit card, can achieve 16TOPS computing performance, can support real-time analysis of 16-channel high-definition video, and is deployed for end-side devices such as cameras and drones, with a power consumption of only about 10W.
Second, the Atlas 300 AI accelerator card: adopts a standard half-height half-length PCIe card design, and is oriented to data center and edge server scenarios. The accelerator card supports multiple data precisions, and a single card can provide 64TOPS INT8 computing performance, providing more powerful computing power for deep learning and reasoning.
Third, the Atlas 500 smart small station: it is an industry-leading edge product that integrates AI processing capabilities. The size of a set-top box can achieve 16-channel high-definition video processing capabilities, which is four times faster than industry products. Smart small stations are applicable to a wide range of fields such as transportation, nursing care, unmanned retail, and smart manufacturing.
Fourth, Atlas 800 AI all-in-one machine: Based on the standard framework and programming environment, it provides an optimized AI environment and pre-installs the underlying software library, which can be used out of the box within 2 hours. At the same time, the AI all-in-one machine integrates Huawei cluster management, task scheduling and other management software and system-level performance monitoring system, which can greatly reduce the threshold for enterprise AI application.
If understood literally, "Atlas" is the giant giant in ancient Greek mythology. This title very aptly describes Huawei's positioning and expectations for the Atlas intelligent computing platform.
In fact, as early as the Huawei Connect Conference in 2017, Huawei released the Atlas intelligent cloud hardware platform for the first time. Based on the FusionServer G series heterogeneous servers, it is mainly oriented to AI, public cloud, and HPC scenarios, and its performance is more than 10 times higher than that of the x86 architecture. Over the past year, Huawei's Atlas intelligent cloud hardware platform has been widely used in smart cities, video recognition and other fields, providing a strong infrastructure support for AI scenarios.
For example, Huawei's Atlas intelligent cloud hardware platform layer has helped Shenzhen traffic police build a traffic brain, enabling second-level search of 1 billion images and real-time analysis by 20,000 cameras, increasing the traffic capacity of Shenzhen's main roads by about 8%, and intelligently assisting law enforcement to increase the efficiency of identifying illegal images by 10 times.
In general, Huawei's Atlas intelligent cloud hardware platform is mainly oriented to the data center side and adopts the industry's mainstream heterogeneous computing architecture. The Huawei Atlas intelligent computing platform released this year is a brand-new evolution and upgrade. Based on Huawei's self-developed Ascend 310 chip, its positioning is more precise, that is, it provides a more targeted and powerful infrastructure for AI applications, and becomes the "Optimus" that supports AI.
The future of Atlas is to sail to the "sea of stars"
In my opinion, Huawei's Atlas intelligent computing platform is of key significance to Huawei's overall AI strategic layout, which can be observed from three dimensions:
First of all, from the perspective of strategic position, Huawei has fully launched its AI strategy, including investing in basic research, building a full-stack solution, investing in an open ecosystem and talent training, enhancing solutions, and improving internal efficiency. The Huawei Atlas intelligent computing platform is not only one of the important achievements of Huawei's basic research, but also the cornerstone of the full-stack solution. It is also a powerful tool for Huawei to build an ecosystem, improve solution capabilities, and internal efficiency. Its strategic position is self-evident.
Secondly, from the perspective of Huawei's goals for AI, Huawei hopes to achieve "inclusive AI". The Atlas intelligent computing platform includes the device side, the edge side, and the data center side, and realizes the collaboration of the device, edge, and cloud, so that intelligent computing can cover the entire scene of AI applications, which truly lowers the threshold for AI applications and makes AI accessible to more users.
Finally, for the entire industry, AI is the core force driving industrial upgrading. The fourth industrial revolution has begun, and we are moving towards an intelligent world. From personal terminals, homes, production and business scenarios, intelligence will be everywhere, and computing will be everywhere. In the process of the entire industry moving towards the "star sea", Huawei's Atlas intelligent computing platform has created new possibilities for customers, allowing customers to flexibly respond to business changes and create business value in the intelligent era.
It is not difficult to foresee that with the advent of the intelligent era, the Atlas intelligent computing platform will play a more critical role for both Huawei and the entire AI industry.
Shenyao's Science and Technology Observation, founded by Shen Yao (WeChat ID: shenyao) , a cross-border technology media person , has 15 years of media work experience, and has self-driving experience of 100,000 kilometers of roads in China and the United States. Laoski has column on major self-media platforms, dedicated to the observation and thinking of the technology industry, here to understand the technology industry, know the trend, and win the future!
