Machine's heart original
Author: Zeannan
directly builds aih framework with chip manufacturers to achieve the highest efficiency.
ai1ai chip can cope with challenges that the general computing architecture cannot cope. However, in order to achieve unprecedented acceleration, we not only need powerful chips, but also deeply intensive fusion optimization with it. Since the outbreak of
deep learning technology, the GPU giant Nivida has established a complete system from chip, system to algorithms and applications to help intelligentization from technology companies to industry, to cutting -edge science and other fields. In China, a company is also developing the "AI operating system" and has gone out of a new model with many hardware manufacturers.
was introduced on the World Artificial Intelligence Conference last week, and Baidu introduced its own software and hardware fusion system.
" Flying Paddy has been pushing AI chip adaptation since 2020. We spent a lot of energy for this. After several years of deep cultivation, we have in -depth cooperation with domestic and foreign chip manufacturers to comprehensively adapt AI chips. Through cooperation, we can truly play the computing power of AI chips, "Ma Yanjun, general manager of Baidu AI Technology Ecological Ecology. "This year, our cooperation with chip manufacturers has entered a new stage of co -creation."
has worked for more than two years. Most of the popular AI chips on the market have obtained the original acceleration of the Baidu fly paddle platform, achieving the industry's leading efficiency of the industry. Essence
High -quality computing power promotes the application of AI technology. These optimized computing power is not only used in the intelligent business of Baidu itself and partners, but also welcomed in the academic and developer groups.
provides native acceleration
for AI chips. It is well known that the rapid development of deep learning is continuously promoting the growth of computing power. Studies have pointed out that with the practicality of deep learning in 2010, the computing power required to train AI is about every 6 months. Since 2015, due to the emergence of large -scale machine learning models, the speed of demand growth suddenly increased to 10 to 100 times per year.
Researcher pointed out that in the past 12 years (2010-2022), machine learning training computing power has increased by 10 billion times.
Index -grade enhanced computing power requirements make chip manufacturers face huge challenges. People try to seek breakthroughs through the AI chip of an innovative architecture. However, on the common frameworks such as pyTorch, although the project is open source as a whole, due to compatibility and other problems, the development team does not accept the code of chip manufacturers to enter the main trunk. The cost of supporting the new version of the framework is very high, and only support to support the key versions.
Baidu flying paddle is committed to incorporating the code of the new AI chip into the main — for more than two years, it has been developing a more convenient core framework with the hardware ecological partner to build a unified hardware access solution.
For the AI framework, every time the code is closed, the model needs to ensure that the model is still stable and correct, which means that a lot of manpower is required to spend a lot of time verification. To this end, the flying paddle actively invited various hardware manufacturers to cooperate to build a verification platform, setting up a special team to test the code for each line of chips. The purpose is only one: to ensure that developers can be convenient to use.
chip running is the basic requirements for use. People can use AI chips to improve efficiency, and fly paddle can also give full play to the design characteristics of the AI chip.
"Each AI chip has its own characteristics. For their special abilities, it can only be realized with the development of a framework with the hardware core research and development team," Ma Yanjun said. After
has cooperated in depth with domestic and foreign manufacturers, the flying paddle framework played the characteristics of the hardware, which can maximize the performance of these chips. In terms of performance optimization, the paddle and NVIDIA are the first to complete the cooperation to support the acceleration capabilities of the structural sparse matrix operation on the NVIDIA Tensor Core.In the training and reasoning tasks of specific machine learning models, the paddle can make full use of the hardware characteristics to greatly increase the operation speed.
HTML on June 30, the latest international authoritative AI benchmark test list on June 30, Baidu uses flying paddle framework and Baidu Smart Cloud 100 舸 舸 calculation platform GPU training performance results. It ranks first in China, surpassing the high -customized and optimized NGC PyTorch framework, which has long been in the leading position of the list, showing the performance advantage of flying paddle to the world.
takes cooperation with NVIDA as an example. Baidu and more manufacturers have started the process of joint research and development. The so -called joint research and development is to jointly polish the basic software stack and promote the adaptation and performance optimization of hardware and paddle. After that, it is the practice and promotion of technology. The technical solution that has been successfully applied to the application will be jointly authorized by both parties and recommend it to ecological partners. In addition, Baidu will also provide development tutorials to bring developers to discuss the latest developments in the AI field with industry experts.
022, flying paddles and Intel1 Intel , Ruixin Micro , ARM, Imagination and other domestic and foreign hardware manufacturers jointly released the "hardware ecological co -creation plan". Basic development stack characteristics, For different application scenarios and products, the customized version of the flying paddle framework is jointly launched, the open source and open model library, the development of courses and training content, etc., and the goal of better service developers.
As of now, 17 member companies have joined the paddle "Hardware Ecological Co -Creation Plan".
This includes star companies in the field of artificial intelligence chips. On Graphcore's dedicated AI chip, the paddle provides perfect support capabilities. For its IPU distributed processor and storage architecture, the flying paddle integrates related interfaces, allowing ordinary developers to fully use the full performance of the chip.
The same thing happens on many domestic chips. Compared with other frameworks, the native of the flying paddle supports more AI chips, which is more convenient and faster for users. "This customization is not visible to upper users." Ma Yanjun said. "The interface used for developers has not changed, but because the deep customization is completed on the frame and chip layer, the performance is optimized to the extreme, and people can experience faster speeds. We solve the framework and chip adaptation The problem, to some extent, also reduces the threshold of user application AI technology. The whole process began to have significant standardization, automation and modular industrial production characteristics, and the threshold is also decreasing. The ability of flying paddle is helping Qianxingbaiye to complete the intelligent upgrade.
In the World Artificial Intelligence Conference "Software and Hard and Hard Emphasized Industry Future" special forum, Intel, NVIDIA, Imagination, Skinhara, Black Sesame Smart, Ziguang Zhanrui , Kunlunxin introduced the results of cooperation with flying paddles, respectively Essence
In the forum, experts discussed the concept of "soft and hard -hard empowerment chip design". As a deep learning platform for open source, flying paddle is a very critical part of the entire industrial chain to undertake AI applications and connect to the bottom. The paddle can quickly pass the development of developers to IP manufacturers at all levels of AI applications such as operators, models, and computing power, and optimize the design of the AI software tool chain from the source of the industrial chain with IP manufacturers. These source work will provide a good foundation for various types of development work in the downstream, and improve the development efficiency of chip design manufacturers and even terminal manufacturers.
For chip manufacturers, the deep support of flying paddle means that the ability of AI chips can be applied by millions of developers. In the perspective of developers, after fully understanding the chip ability, how to choose the right AI chip for your own work is no longer a complex problem.
uses a large number of different AI computing power in the ecosystem of paddle services from smart clouds and C -end business to the end -side AI and IoT devices of the service industry. Different types of chips can find places to play value. As of now, as of now, the number of domestic and foreign hardware manufacturers who have cooperated with flying paddle has exceeded 30 domestic and foreign hardware manufacturers. The mainstream machine learning chips at home and abroad have basically adapted to the paddle. Baidu's use of products with chip manufacturers has made many different AI chips find widespread application scenarios.
can promote ecological construction only with cooperation with more open and truly commercial value. The exploration of paddles in the coordinated hardware and hardware has been positioned for the application of the leading AI framework.
reduces AI model thresholds, helping developers
is worth mentioning that the ability of flying paddle provides not only is widely used in the industrial industry, but also welcomed in the academic and developer groups.
Baidu provides AI technology on the one hand, and is also a large -scale user of AI computing power. Inside the company, the "Baiyao" AI heterogeneous computing platform must run 180,000 training tasks per month, and the AI model of ordinary users must call each search every time, and 6 billion requests are required every day.
These needs test the real -time response capabilities of AI infrastructure. "After Baidu Smart Cloud's in -depth and intellectual transformation and upgrading of the industry, the demand for AI chips has changed. Business needs, "Ma Yanjun said. "In some Baidu's business, the big model has become part of the workflow. The practice of" Wenxin "is real, as long as you call the interface, you can use it."
said that the big model is when it comes to the big model. "Can't afford it." However, flying paddle has been constantly reducing the threshold of large model applications, supporting the large -scale production and industrial application of large models from the training, reasoning, and compression of large models.
, especially for the support of the academic community, flying paddle has been providing AI computing resources. According to reports, 70% of the undergraduate colleges and universities opened in Shanghai are teaching with flying paddle, including Shanghai Jiaotong University 's artificial intelligence programming practice, Fudan University machine learning, Tongji University computer science introduction, etc. course.
In the AI course of the university, the flying paddle provides free computing power and a large number of teaching content. Even the ability of the pre -training large models can be called "three -line code" through Paddlehub.
"After we open a large model on PaddleHub, because the user's usage is unexpectedly increased, the background server quickly crowded." Ma Yanjun said. "Professor and students are objective and rational groups. Only when you do something well can people really use it."
deep learning framework is considered "operating system in the intelligent era" as the number one in deep learning in China Framework and empowerment platform, the world's top three artificial intelligence and open source open ecology, flying paddle has a core framework with flexible, efficient and extensive adaptation, rich functional and widely scenarios, and more and more industry development. Those who are playing new productive forces in their ecology. As of May this year,
has attracted 4.77 million developers, serving 180,000 enterprises in industrial applications, and more than 560,000 AI models have been applied to the platform.
, as Baidu CTO Wang Haifeng said: "Based on the flying paddle platform, everyone can become developers of AI applications."
continues to land, hardware computing power and software algorithm will enter the collaborative innovation of collaborative innovation. In the new stage, a large wave of cooperation in flying paddle has taken an important step on the road of "soft and hard work".
Reference content:
https: //github.com/ml-progress/compute-trends
