Huawei Zhang Dixuan: Jointly Building an Intelligent Foundation and Leading a New Future of AI

Tech 2023-07-14 09:39:31 Source: Network

On July 6th, the Shengteng Artificial Intelligence Industry Summit Forum was held in Shanghai. At the meeting, Zhang Dixuan, President of Huawei's Ascension Computing Business, delivered a keynote speech titled "Building a Smart Foundation Together and Leading the New Future of AI"

On July 6th, the Shengteng Artificial Intelligence Industry Summit Forum was held in Shanghai. At the meeting, Zhang Dixuan, President of Huawei's Ascension Computing Business, delivered a keynote speech titled "Building a Smart Foundation Together and Leading the New Future of AI".

He said that only the towering trees of the artificial intelligence industry with a solid foundation, deep technology, and closer ecology can prosper and develop. Huawei Shengteng will firmly invest in the field of artificial intelligence, lay a solid foundation for software and hardware computing power, and jointly create a new era of artificial intelligence with partners and customers.

The following is the full text of Zhang Dixuan's speech

Dear leaders, experts, and guests, good afternoon everyone!

Today, it is a great pleasure to gather with you again at the WAIC Ascension Artificial Intelligence Industry Summit. In Shanghai, a city with a profound cultural foundation and rapidly changing development speed, we have witnessed the rapid development of the Ascension AI industry ecosystem together. On behalf of Huawei's Ascension computing business, I would like to welcome everyone again!

The theme of my speech today is "Building Intelligent Foundations Together, Leading the New Future of AI".

In the past year, the development of the Shengteng AI industry has accelerated, entering more industries, cities, universities, and research institutions. More partners and developers have chosen Shengteng as their first choice for innovation.

As of now, Shengteng has been certifiedthirty manyHome hardware partnersone thousand and two hundred manyHome software partners, jointly incubated2500many AI2AI1manythirty0many100000Professional AI talents have achieved rapid growth and doubled the number of developers, surpassing1 million 800 thousandPeople. As the main direction of industrial development, large models,2020thirtymany

The development of the Shengteng industry cannot be separated from the joint efforts of every industry colleague. Once again, I would like to express my gratitude to leaders, experts, partners, and developers at all levels for their trust and support in Shengteng AI along the way.

I believe everyone shares the same feeling. With the support of massive AI computing power and innovative algorithms, artificial intelligence is rapidly developing, further affecting and even changing our world. In three words, it can be summarized as: crossing, integrating, and creating.

The first step is to cross:With the successful commercialization and implementation of large models represented by ChatGPT, it has caused industry shock and pursuit. Just as visual technology has crossed a turning point in 12 years, AI technology has now moved from perceptual understanding to the stage of generative creation, which is a key milestone.

Next is fusion:The powerful generative and creative capabilities have also enabled artificial intelligence to truly integrate with our production, daily life, and even scientific research, moving from auxiliary intelligence to native intelligence. At the beginning of design, various industries have regarded AI as the core element.

Finally, creating:When the marginal cost of using any general-purpose technology gradually decreases, it will create far greater social value and economic benefits than expected. It is predicted that generative artificial intelligence will create trillion level digital economic growth, from supporting acceleration to becoming the core engine of economic growth now.

Fortunately, we are all in this golden age of AI explosion, undertaking different missions.

Since Huawei released its AI strategy in 2018, it has persisted in innovation around root technology and focused on building basic AI capabilities. It has created a "ultimate performance, extremely easy to use" Ascension AI basic software and hardware platform, and comprehensively upgraded its training and reasoning capabilities based on industry scenarios. At the same time, we are also following the industry's suggestions and feedback, continuously improving our capabilities in operator development, dynamic shape, and open compatibility, and have achieved phased results.

5Ascend CCC++22manyKernelShape95%CVOCR

In terms of open compatibility, providing pluggable adapters can better support industry frameworks, and the adaptation of new versions can be completed in one month, allowing everyone to quickly innovate and practice on Ascension based on the existing ecosystem.

many

In order to better promote the development of artificial intelligence, we continue to lead and innovate in three major areas, accelerating towards the new era of AI:

Firstly,Leading innovation in computing powerThrough system level innovation, from single point computing power to cluster computing power, development has a foundation.

Secondly,Breakthrough in Software architectureThrough soft and hard collaborative innovation, the entire process accelerates the development of large models, providing a source of innovation.

Thirdly,Enabling Industry UpgradeChange development models, promote industrial collaboration, bridge technology to industry application breakpoints, and enable applications to scale.

7503GAPICTthirtymanyAI

Firstly, through the overall design of the system, we have changed the simple stacking of servers in the past, integrating computing, storage, network, energy, etc., which is equivalent to turning an AI data center into a supercomputer to achieve a doubling of energy efficiency.

Secondly, based on the backplane bus architecture design, it achieves full node blind insertion and precise cooling with full liquid cooling, with a higher computing power density and a PUE lower than 1.15, making the computing power center green and energy-saving, and enabling more flexible deployment and expansion.

many

Altas 900 AI098%194816AI161750100Bthirty

AI AIAImany

Based on Shengteng AI, we delivered Pengcheng Cloud Brain's first 1000P AI computing power cluster in 2020. As of now, it has supported 25 cities across the country, focusing on government, industry, academia, research and application, and has delivered a cumulative computing power scale of 5000P. Among them, 6 cities have achieved computing power saturation operation and expanded their capacity to better support the rapid development of the local artificial intelligence industry. Recently,There are 7 artificial intelligence computing centers based on Ascension, including Wuhan, Beijing, Xi'an, Chengdu, Dalian, Shenyang, etc., which have received national recognition and become one of the first batch of "National New Generation Artificial Intelligence Public Computing Power Open Innovation Platforms" by the Ministry of Science and Technology.

We will replicate the accumulated experience in AI cluster construction in cities to the industry to better support the construction of industry clusters such as operators, finance, and the internet, enabling innovation in industry models.

510thirtymanyGPTGLMLLAMANLPmanymanyShengteng AI supports nearly half of China's large model native innovation.Shengteng is also the only Technology roadmap to complete the development of the 100 billion parameter large model in China at present. The actual training performance of the open source Transformer type large models has achieved 1.2 times of that of the industry,It has become one of the preferred routes for big model innovation.

Today, under the guidance of the Ministry of Industry and Information Technology and the China Electricity Standards Association, we have jointly established the "Large Model Industry Application Working Group", believing that it will definitely better promote the implementation of large model applications in China.

In order to better support and accelerate the innovation of the big model, we also continue to work hard on the basic software such as CANN, framework, Toolchain, and enable the big model in the whole process around development, training, fine-tuning, reasoning and other links.

API-based-basedmany

manyAscend Speedmany15%

manyLoRAP-Tuning5%Gitte

The large model has a huge number of parameters, which makes it difficult to deploy reasoning and high costs. Therefore, we have integrated automatic pruning, distillation, and quantization tools on MindStudio to reduce accuracy loss<0.5%20

Finally, in the inference stage, based on the Ascension inference engine, it can interface with the inference APIs of various frameworks and business middleware systems, support online distributed inference, and achieve rapid application launch with inference delay<50ms

Operators and models are the core assets of users, and Ascension is open through the capabilities of basic software, fully compatible with the industry's tripartite ecosystem, and supports efficient migration and adaptation.

The migration development for third-party operators, combined with AscendC programming language and MindStudio operator migration tool, can achieve GPU operator migration and debugging within 2 weeks.

For model development, we are also compatible with the third-party ecosystem on PyTorch, such as the model development kits of communities such as OpenMMLab and HuggingFace, as well as various acceleration libraries such as Deepspeed and Megatron. Without additional code, we can directly refer to configuration files and implement development based on the Shengteng+PyTorch model in a componentized manner.

Finally, for large models incubated based on GPUs in the industry, Ascent also provides a one-stop tool to support efficient and lossless migration of large models to Ascent. With two lines of code, script analysis and migration can be completed at the minute level, and precision optimization of the whole network model and positioning processing of training performance can be completed at the sky level. The end-to-end lossless migration cycle of the large model has been greatly improved from the original monthly level to about two weeks now, and has been fully recognized by customers such as IFlytek.

We have great computing power and big models, and how to get through the last mile of industry landing is something we need to solve. Especially in the era of large models, there are a series of challenges such as long development cycles, high deployment thresholds, and customer environmental requirements. The traditional development model of accumulating machines and testing is no longer able to meet the needs, resulting in the inability to scale and implement.

To this end, we have collaborated with upstream and downstream partners to reshape the development process for large model solutions, changing the original serial development mode of hardware computing power and model algorithm development. We work with our partners to jointly define the market, jointly innovate, and jointly market, shorten the R&D process, and improve the Product lifecycle.

AImanymany

manyAImany

At present, AI has demonstrated its capabilities in molecular simulation, fluid simulation, electromagnetic simulation, weather forecasting, and other fields, demonstrating exponential performance improvement. Just in the morning, the "Dongfang Wingfind" large model developed based on Ascension also won the highest honor award at the WAIC conference - SAIL Award, which reduced the time required for wing aerodynamic simulation to one thousandth of the original time. It will also accelerate leapfrog innovation in the field of aerodynamics in the national economy and people's livelihood industry.

AImany

The development of the Ascending Ecology is also a process of growing and winning together with partners. We will firmly develop a business ecosystem based on enabling partners.

manymanymany1211IHV

APNmanyAPN12APNmany

Technology is the root, talent is the foundation. The development of the Ascension industry has always placed talent cultivation at the core.

For college teachers and students manythirty0many2000manythirty00many16manyAI

Targeting scientific research talentsRelying on the Shengteng Scientific Research Innovation Enabling Program, we provide funding, resources, collaborative innovation, and influence building support for exploring cutting-edge fields, supporting the cultivation of composite high-end scientific research talents.

For developersBased on the three major platforms of Shengteng Community, Chuangxiangri City Tour, and Shengteng AI Innovation Competition, we have provided a complete "learning training competition" growth system for developers. In the next three years, our goal is to develop an additional 3.5 million Shengteng developers. We look forward to everyone joining us, and together we gather the spark of wisdom to provide a source of sustainable development for the artificial intelligence industry.

many

Let AI innovation "cluster" and AI accelerate "in the future!

Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])

Mobile advertising space rental

Tag: Huawei Zhang Dixuan Jointly Building an Intelligent Foundation and