Artificial intelligence is moving towards the stage of scale, but the breakthrough and application of artificial intelligence technology is facing huge challenges. Long training time, scarcity and expensive computing power, high development costs, and high technical thresholds are the current situation of the industry faced by every developer, and they are also problems that hinder the further development of the AI ecosystem of the whole industry.
Under such a dilemma, Ascend launched an AI-based software and hardware platform, and through a series of landing activities, more AI developers benefited. On April 23, the first stop, Ascend AI Developers Creation day activity, was held in Xi'an, with the theme of "Creating the Future, Enjoying extraordinary", interpreting the value framework of Ascend AI developers. By participating in this activity, developers will get a glimpse of the future development trend of AI, acquire cutting-edge AI technology, update their knowledge system, and add weight to the workplace.
1
"Hand workshop" development model Out, another can be good?
The current AI development model faces inefficient bottlenecks. Over-reliance on experts and data has led to AI often falling into a situation where there is no "intelligence" without "human" and no "intelligence". In addition, in the face of massive data and large-scale pre-trained models, computing power is scarce and expensive, which is also a challenge that developers and enterprises have to face. In particular, the trend of full-scenario layout of terminal, edge, and cloud has deepened the complexity of AI technology, and developers need to carry out multiple development and deployment to achieve full-scenario adaptation, which further raises the threshold of development. How to break the computing power gap, make AI inclusive, and become a real productivity is an urgent problem to be solved in the current industry.
As the computing industry enters the era of intelligence, domestic and foreign giants have accelerated the layout of AI. In contrast, foreign companies in the field of artificial intelligence earlier and more mature, while domestic enterprises are relatively young, in the early days more focused on the AI application layer layout, rarely involving the underlying technology. However, the lack of basic software and hardware support will make it difficult to promote the landing of the industry, and in recent years, domestic companies have also realized this problem.
Based on years of research and accumulation, Huawei innovates the Ascend AI basic software and hardware platform independently, providing Atlas series hardware, heterogeneous computing architecture CANN, all-scenario AI framework Ascend MindSpore, Ascend Application Enabling MindX, and one-stop AI development platform ModelArts, etc., to help partners and developers efficiently use AI capabilities, innovate scenario-based AI applications, and accelerate intelligent upgrades in thousands of industries. On this Ascend AI Developer Creation Day, technical experts from Huawei carefully interpreted the latest technology of Ascend AI.
At the level of AI framework, Ascend MindSpore previously released a new version of 1.6, bringing new upgrades in terms of ease of use, development efficiency, control flow performance and so on. Yu Fan, software scientist, deputy chief expert, and ascendant MindSpore architect of Huawei 2012 Laboratory Central Software Institute summarized the features of the new version 1.6 into the following eight points:
Yu Pan, software scientist, deputy chief expert, and architect of MindSpore, Huawei 2012 Laboratory Central Software Institute
Extreme heterogeneous parallelism, supporting stand-alone 8 card 100 billion parameter model training. Memory Swap can stage part of the data on the host side, helping users train larger-scale models without adding devices, and solving the scenario that GPU/NPU memory is not enough due to the excessive scale of the model during training;
Development kit MindSpore Dev ToolKit provides run management, intelligent knowledge search and intelligent code completion functions to solve the accuracy and efficiency problems of dynamic language completion through unified modeling;
MindSpore1.6 adds support for custom AICPU type operators, which can be quickly deployed to mainstream embedded platforms, and has greater performance benefits for some operators that are difficult to vectorize;
Provide high-performance, flexible and easy-to-use reinforcement learning frameworks, give full play to the heterogeneous computing power of equipment, and support large-scale asynchronous training;
The enhancement of mindSpore Lite inference supports openGL texture data as input and output, which reduces the data copy between CPU and GPU during the end-to-end inference process, so as to improve inference performance and reduce power consumption.
Improve inference and fine-tune efficiency by rapidly equivalently migrating model structure and weight information through the model migration tool MindConvert.
MindQuantum 0.5, which uses quantum neural networks to solve the combinatorial optimization problem MaxCut.
MindSpore Graph Learning provides an efficient and easy-to-use graph learning framework that supports rapid iteration of graph neural networks in application scenarios; formulas as code accelerate training by 3 to 4x.
Through multi-dimensional innovation, The Ascend MindSpore 1.6 version has continuously upgraded the developer experience and has made breakthroughs in many scientific fields. According to Yu Fan, MindSpore version 1.7 will also be released soon, and the new version will jointly release industry-leading results in conjunction with well-known research institutes for protein structure prediction and folding problems.
At the software level, CANN 5.0 fully releases the surging computing power of the original Da Vinci architecture through deep co-optimization of software and hardware, and achieves a significant improvement in AI model training / inference performance in terms of performance. At the first stop, Ascend AI Developer Creation Day, Song Kunpeng, AI computing architect of Huawei Ascend, introduced the key breakthroughs achieved by CANN 5.0.
Song Kunpeng, AI computing architect of Huawei Ascend
COMPARED TO VERSION 3.0, CANN 5.0 doubles performance in large-scale cluster training and common model training, and can achieve a 30% to 140% improvement in typical inference scenarios. Its improvement comes from the breakthrough of many key technologies, such as automatic flow optimization, deep integration of operators, adaptive gradient slicing, AOE intelligent calculation tuning and so on.
For oversized model training, CANN5.0 relies on automatic decomposition and parallel technology to make it as simple as the processing of the normal model, and solves the processing problem of large format and multi-channel remote sensing image.
For reasoning services, CANN5.0 accelerates the development and execution of inference services by relying on the inference toolbox, helping developers achieve high-performance inference applications:
For model compression requirements, CANN5.0 provides the model compression tool AMCT, which supports quantization, channel sparsity, and tensor decomposition, reduces the amount of data and computation of the model, and improves performance.
For model performance tuning, CANN 5.0 provides AOE, a ascend tuning engine for customer networks, which supports the automatic optimization of the operator calculation process and improves the computing performance of the entire network.
In terms of development and deployment, CANN 5.0 also supports online inference and offline inference to meet the deployment requirements of different scenarios, and developers can choose according to different scenarios.
In summary, Ascend AI continues to empower developers through new technologies, lower the threshold for development, and accelerate the deep integration of AI and industry. As Jiao Licheng, dean of the Artificial Intelligence Research Institute of Xidian University, foreign academician of the European Academy of Sciences, and foreign academician of the Russian Academy of Natural Sciences, said: "The development of artificial intelligence is inseparable from the support of core technologies, these core technologies are 'card neck' technologies, without these technological breakthroughs, it is difficult to promote the development of society, and the mission of Ascend AI is to solve problems that others cannot solve." ”
2
Open up the last mile of AI industry application landing
The development of technology must eventually be deeply integrated with the industry. Today, AI is not just a technology, but a new productivity. With the acceleration of intelligent upgrading, in smart cities, manufacturing, transportation, energy, finance, education and other industries, Ascend AI relies on technological innovation to continue to deepen. Ascend AI developers choose Sanqin Fertile Land Xi'an for the first time, and Shengteng AI has a wealth of innovation-driven cases in Xi'an.
The ARTIFICIAL COMPUTING CENTER IS THE INFRASTRUCTURE FOR THE CITY'S PUBLIC COMPUTING POWER. Xi'an Future Artificial Intelligence Computing Center, which is built based on Ascend AI's basic hardware and software platform, serves as the artificial intelligence industry innovation base of Qinchuangyuan Innovation-driven platform in Shaanxi Province, providing inclusive AI computing power for universities, scientific research and enterprises. At present, Xi'an Future Artificial Intelligence Computing Center has cooperated with 112 colleges and universities and enterprises, involving remote sensing, cultural tourism, manufacturing, power and other industries, providing local computing services for developers.
At the level of industrial applications, Ascend AI has a wide range of applications. Taking safe production as an example, whether it is aerospace, energy, or chemical industry, a large number of scenarios also use the traditional manpower group construction method to solve the problem of supervision in safe production, which can only be traced back to the source after the occurrence of accidents or problems, and cannot achieve advance warning or real-time online supervision on the spot. In view of the above pain points of production and innovation enterprises, Xi'an Xiangxun Technology and Ascend AI have developed an intelligent supervision system to achieve pre-warning, in-process warning, and post-event traceability. The program realizes the information collection, intelligent identification of hidden dangers, capture of violations, safety warning, statistics and decision-making in the production workshop and dangerous environment, and solves the problems of difficult to assess the compliance and management status of the factory and workshop, and the difficulty of accurate scientific decision-making. Provide an accurate basis for enterprises to assess the management status of safety production and adopt targeted improvement methods.
In terms of scientific research and innovation, Ascend AI combines with universities to promote the integration of production, education and research, and help the development of scientific research. Relying on the Xi'an Future Artificial Intelligence Computing Center, Xidian University conducted research in the field of radar and remote sensing, created the world's first radar remote sensing pre-training model, released the world's first radar remote sensing development kit, and also released the world's first radar remote sensing sequence memory and the world's first radar remote sensing framework features. In addition, Ascend AI also has in-depth discussions and cooperation with Northwestern Polytechnical University in speech pre-training big model technology, and has joined hands with Shaanxi Normal University to build MindSpore Research Laboratory to promote innovation, achieve the transformation of artificial intelligence achievements, and achieve the goals of production, learning, research and use.
Artificial intelligence is playing a huge value at the industrial level and scientific research level, but for ordinary people, how to get closer to artificial intelligence? Zhihui Jun (Peng Zhihui), who is known as "Wild Iron Man" on Bilibili, has used Ascend technology to complete the development and landing of a variety of robots, such as parts grinding robot arms, tree planting robots, etc. He said: "For humans to build a robot is not the biggest difficulty, the more difficult thing is how to make the machine have the same intelligence as a human." Based on the support of ascend coverage scene and the components provided, it can not only meet the requirements of application computing power, but also ensure the reliability of chips and hardware. ”
3
Co-creation of the ecosystem, this is the best era for developers
Technology is the core, industrialization is the goal, but if you want to achieve continuous innovation and development, ecological construction is indispensable. In order to develop the ecology and cultivate artificial intelligence talents, Ascend AI has launched the Crowd Intelligence Program since 2020, and as of now, more than 2,500 developers have participated in it and more than 2,800 operator models have been built.
Tian Shuai, a master's student at Xidian University, and his team participated in the process of co-construction of Ascend Ecology. From contact with operator analysis so far, they have participated in a total of 50 operator analysis, 15 operator development, 26 Ascend MindX model development, 3 TensorFlow model development, involving 13 scenarios of 18 models, some of which are output in the form of enterprise-level solutions, he hopes that these models can finally be truly landed and used to solve problems in production and life.
Zhang Dixuan, president of Huawei's Ascend Computing Business, said that The Ascend Wisdom Plan 2022 plans to invest 200 million yuan in incentive funds, and is expected to launch More than 3,200 Operator Models, More Than 400 Application Cases, and More Than 50 Innovative Applications developed by Smart Wisdom.
Ascend AI has always embraced developers across the country with an open attitude, and Ascend AI Developers Create a Day, which is a series of activities launched by Huawei for ascend AI industry full-stack developers. Through live sharing, interactive discussions, immersive hands-on experience, etc., developers can get a close look at the latest software, tools, technologies and practices of the Ascend AI industry, update their knowledge system, broaden their horizons, and improve their capabilities.
On the Ascend AI Developer Creation Day, the first Ascend AI Innovation Competition was also officially launched. The competition will gather developers from all over the country to communicate with each other, challenge cutting-edge innovative technologies, solve problems, and achieve personal breakthroughs. The competition has 15 sub-divisions, including the Application Track and the Ascend Track, with a prize pool of up to 10 million, and interested developers can sign up to participate in the challenge.
In the previous Internet + Innovation and Entrepreneurship Competition, Chen Tianyi, a doctor from South China University of Technology, won the gold medal of the competition based on the AI framework Ascend MindSpore. As for why he chose MindSpore, he said: "On the one hand, MindSpore has established a friendly API that developers can get started with; on the other hand, MindSpore can greatly improve development efficiency in data computing, data processing, and distribution; third, MindSpore supports cloud, edge, and end-side scenarios, realizing full-scenario development, and more widely used and convenient; in addition, Ascend has improved rich learning resources for MindSpore. Beginners can also get started quickly. ”
In ascending AI, there is a popular saying that "the strongest wisdom is the wisdom of the crowd, and the greatest force is the joint force". By building an expert growth system, Ascend AI provides a growth channel for developer groups who are interested in research and innovation in the field of Ascend AI, and sets up a corresponding rights and interest system to motivate experts and encourage experts to exert influence and creativity.
Scan the code to register
Click on one to see fewer bugs