Home >Technology peripherals >AI >The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

PHPz
PHPzforward
2023-04-09 09:41:021667browse

On September 13th, China’s only feast focusing on autonomous driving AI technology, the sixth HAOMO AI DAY golden autumn is coming. The theme of this year is "A journey of wisdom starts from the beginning". HAOMO AI DAY, which has always had a strong technical background, still brings the most advanced technological achievements and product releases this time. It was the 1020th day since the founding of HaoMo. HaoMo announced its latest report card for the fall of 2022:

HaoMo has reached the fastest 1,000 days of autonomous driving in China, ranking first in mass-produced autonomous driving in China. , has steadily delivered three generations of passenger car assisted driving products in two and a half years, and has now been equipped with more than ten star models. China's first large-scale mass-produced urban NOH model, the Mocha DHT-PHEV lidar version, is scheduled to be mass-produced in September and will be released within the year. . The Wei brand Mocha PHEV and Euler Haomao equipped with Feimo HPilot have obtained the EU E-NCAP "five-star safety rating", making Feimo the first Chinese autonomous driving company to go into mass production overseas. In terms of terminal logistics automatic distribution, it occupies the absolute leading market share in this field. Xiaomotuo 2.0 has been mass-produced and delivered to customers. MANA, China's first data intelligence system pioneered by Haomo, has completed the annotation of hundreds of thousands of all-factor, multi-modal CLIPS, accumulated 3 million hours of Chinese road driving cognitive scene libraries, which is equivalent to 40,000 years of human drivers, and has basically completed the data closed loop...

Zhang Kai, chairman of Haomo Zhixing, said: "At the end of 2021, we expected that 2022 would be an extremely difficult year. But after experiencing the epidemic and various black swan events in the first half of the year, we found that the actual situation It is more difficult than the prediction at the end of last year. Through the hard work and unremitting efforts of the Weimo people, we are overcoming the many challenges brought by the epidemic, the supply chain, and the technology itself. Facing the future, Weimo is full of confidence, and will focus on mass production next Deliver answers to the big test of large-scale intelligent driving and move towards the next 1,000 days of Hao Mo.”

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(Picture: Zhang Kai, Chairman of Hao Mo Zhixing, delivering a speech)

The exclusive secret of Haomo’s 1000 days revealed: The Iron Triangle of Intelligent Driving Product Iteration creates a different Haomo

In the sixth HAOMO AI DAY speech, Zhang Kai spoke of "Haomo's 1000 Days: With the title of "New Cycle, New Journey", it reviewed the growth process and experience of Haimo since its establishment and has successfully crossed the line between life and death of startup companies and become a new benchmark for Chinese autonomous driving technology startups. It also shared its judgment on the macro trend of intelligent driving and Haimao's The latest progress of the three major battles in 2022.

It’s not easy to start a business. How did Wei Mo win, successfully create a unique “Hai Mo model” and technology moat, and significantly lead the competition? In his speech, Zhang Kai attributed this to Haimo's correct choices and continuous efforts in strategic routes, business and technological breakthroughs, and product iteration capabilities.

"HaoMo Zhixing has always been firmly following the path of progressive development. In the era of autonomous driving 3.0, assisted driving is the only way to autonomous driving." Zhang Kai addressed the above issues at the beginning of his speech The first answer was thrown out. At present, China has become the main battlefield for smart cars in the world. It is expected that by 2025, the installation rate of high-level assisted driving will exceed 70%. Under the huge trend of the era of intelligent driving, "data is the core essence of driving the maturity of autonomous driving, and the progressive route is the best path for data accumulation" is increasingly recognized by the industry and has become a common development direction of autonomous driving companies. By virtue of its firm implementation of the progressive route, Bomo has successfully explored the unique "Haimo Model", which has become an important direction and cornerstone of Bomo's 1,000-day rapid development.

The strategic line serves as the direction guide, and breakthroughs at the technical and commercial levels are the key to victory. “In the past 1,000 days, Wei Mo has bravely ventured into the no-man’s land of the industry and overcome the three major difficulties in the industry from 0 to 1—the difficulty of mass production of large-scale, multi-model autonomous driving, the technical cost difficulty of end-logistics automatic delivery vehicles, and the difficulty of large-scale data processing and There are difficulties in applying large models." Zhang Kai also pointed out that breakthroughs in the three major difficulties, coupled with efforts in ecological construction, security, culture, talent and other aspects, support Weimo's ability to cross the life-and-death line of startup companies and achieve rapid development.

"HaiMo Zhixing has become the earliest and only autonomous driving company in China to enter the rapid product iteration stage." Zhang Kai said that HaoMo's polishing of product iteration capabilities has created a different BiMo and become HaoMo's 1,000-day company. Tips for winning. According to Zhang Kai, Feimo has now created a unique "iron triangle of intelligent driving product capability iteration", in which scenario-based user experience design is the entrance, artificial intelligence technology is the soul, and technical engineering capabilities are the guarantee. There is an entrance, a soul, and a guarantee. The three major capabilities are in a positive cycle and support each other, allowing Haimo's intelligent driving product capabilities to be rapidly iterated. Haimou Zhixing is the company with the most experience in mass production and implementation of intelligent driving technology in China. Up to now, more than ten passenger models equipped with Haimou HPilot products have been put into mass production. Models already equipped with them include Wei Brand Mocha and Wei Brand Latte. , Wei Brand Macchiato, Tank 300, Tank 500, Haval Beast, Latte DHT-PHEV, etc. Models currently being delivered include Mocha DHT-PHEV lidar version, Euler Lightning Cat, Euler Ballet Cat, and a new generation of Great Wall Cannon. It is also the only company in China that has the ability to develop more than 30 intelligent driving projects asynchronously and in parallel. In terms of process development of intelligent driving products, matching was completed within 2 months, and the calibration efficiency ranks first in the industry; intelligent driving products have been robustly designed to achieve a 100% one-time pass rate.

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(HaiMo Intelligent Driving Product Capability Iteration Iron Triangle)

More than 1,000 days and nights of risk-taking, HaoMo Zhixing created China’s autonomous driving market with its “HaiMo Model” Driving a new paradigm. Standing at the 1,000-day milestone, the three major battle flags set by Wei Mo are also entering a critical period for victory. So what strategy will Wei Mo use to advance? Zhang Kai announced the "Five Winning Rules for Winning the Second Half of Intelligent Driving" at the scene: Intelligent driving product development always puts safety first; the product experience is "really fragrant"; it is based on the user's real scene data drive to achieve rapid product iteration; Perceptual intelligence and cognitive intelligence are highly integrated; empower customers with an open mind and promote common progress in the industry.

“Many great projects are only completed in the second half of the year. In the next four months, we will make every effort to sprint towards the mass production delivery target set at the beginning of the year. We believe that the dawn of victory is ahead! "Zhang Kai said.

Sprinting into the era of autonomous driving 3.0, MANA ushered in six major milestone upgrades to protect urban NOH

Grasp the cutting-edge development trend of AI and maintain its technological leadership Safety has always been the criterion for developing autonomous driving technology. At this event, Dr. Gu Weihao, CEO of BoMo Zhixing, shared his insights into the current development trends of autonomous driving technology with the theme of "HaoMo and the 3.0 Era of Autonomous Driving", and proposed for the first time in the industry: Autonomous driving has entered the The data-driven 3.0 era.

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(Speech by Dr. Gu Weihao, CEO of HaoMo Zhixing)

He pointed out in his speech that the Attention large model is a new trend in the current development of AI, but due to Its demand for computing power far exceeds Moore's Law, and the resulting challenges such as high computing power demand, high training costs, and high implementation difficulty are becoming the current key issues to be solved. "Haimo is reducing the cost of autonomous driving through low-carbon supercomputing, implementing large models on the vehicle side by improving the design of car-side models and chips, and making large models more effective through data organization." Gu Weihao said, Based on the Attention large model, autonomous driving requires large-scale and diverse training data. Only passenger car assisted driving based on large-scale real human driving data can accumulate sufficient scale and diversity of data. "We have reason to believe that assisted driving is the only way to autonomous driving. Because only assisted driving has the ability to accumulate data of sufficient scale and variety."

“The era of data-driven autonomous driving 3.0 has arrived.” Gu Weihao said that the development of autonomous driving in the past ten years can be divided into three eras: hardware-driven, software-driven, and data-driven. The data-driven era is a completely different era. Large models and massive data are a "double sword", and the data turns on the self-training mode; in perception technology, multi-modal sensors are used to jointly output results; in cognitive technology, interpretable scenarios are used Common sense of driving is the main focus; autonomous driving mileage is driven by hardware and software, and the mileage has soared from millions to tens of millions of kilometers in the era of hardware and software to more than 100 million kilometers. "HaiMo has been preparing for the era of autonomous driving 3.0. In terms of perception, cognition, and model construction, it is all built in a data-driven way. Everything HaoMo does is to be able to build data channels and computing centers. In order to obtain data more efficiently and transform data into knowledge." With data drive as the core, only when the above four technical conditions are established in parallel can we truly enter the era of autonomous driving 3.0. Currently, Tesla has led the world in entering the era of autonomous driving 3.0, and Haomo is most likely to become the first Chinese company to enter the era of autonomous driving 3.0.

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(Illustration: Feimo Zhixing sprints into the era of autonomous driving 3.0)

The ultimate pursuit of leading technology not only allows Feimo to always be at the forefront of industry innovation , while also promoting rapid iteration of its own products. According to Gu Weihao, Haomo started research and implementation attempts on the transformer large model as early as June 2021. It is based on the successful practice in the past year and more in the transformation and upgrading of training platforms, preparations for switching data specifications and annotation methods, and exploration of model details for specific tasks of perception and cognition, etc., that we have laid the foundation for the current situation of Hao Mo in urban navigation assisted driving scenarios. Rapid development has laid a solid foundation.

In addition, Gu Weihao also shared the challenges and major functional upgrades faced by the MANA data intelligence system in autonomous driving urban scenarios.

Gu Weihao said that urban roads mainly present “four types of scene problems and six major technical challenges.” The scene problems mainly include "frequent urban road maintenance", "dense large vehicles", "narrow lane changing space" and "diverse urban environment". To solve the above scene problems, we face six major challenges at the technical level: how to apply large models in the field of autonomous driving, how to make data play a greater value, how to use re-sensing technology to solve the problem of real space understanding, how to use the interactive interface of the human world, how to To make the simulation more realistic, how to make the autonomous driving system move more like a human.

In order to cope with the above challenges, MANA’s sensory intelligence, cognitive intelligence and other aspects have been updated and upgraded.

First of all, MANA uses a self-supervised learning method using unlabeled data from large-scale production vehicles to create model effects. Compared with training with only a small number of labeled samples, the training effect is improved by more than 3 times, which makes Haomo Data advantages can be efficiently transformed into model effects to better adapt to the needs of various perception tasks for autonomous driving.

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(Illustration: MANA turns on self-supervised learning)

Secondly, MANA’s perception capability has been improved, so that massive data will no longer be treated differently. Faced with the problem of "data efficiency" under huge data scale, MANA built an incremental learning and training platform, extracting part of the existing data and adding new data to form a hybrid data set. During training, the output of the new model and the old model are required to be as consistent as possible, and the fit to the new data is as good as possible. Compared with conventional methods, the overall computing power is saved by 80%, and the response speed is increased by 6 times.

Third, MANA has stronger perception ability. By using the time-series transformer model to perform virtual real-time mapping in the BEV space, the output of perceived lane lines is more accurate and stable, allowing autonomous driving in urban navigation to bid farewell to high-precision maps.

Fourth, MANA’s perception capability is more accurate, so there are no vehicle signals that cannot be recognized in China. By upgrading the on-board perception system, MANA specifically identifies the status of brake lights and turn signals, allowing drivers to be safer and more comfortable in scenarios such as sudden braking or emergency cuts of the vehicle in front.

Fifth, MANA’s cognitive ability has also evolved again. Facing the intersection, the most complex scene in the city, MANA introduced high-value real traffic flow scenes into the simulation system. It cooperated with Zhejiang Deqing and Alibaba Cloud to introduce the intersection, the most complex scene in the city, into the simulation engine to build an autonomous driving scene library. , through the real simulation verification of autonomous driving, the timeliness is higher and the microscopic traffic flow is more realistic, effectively solving the "old difficulty" problem of passing through urban intersections.

At the event site, Haomo, together with Zhejiang Deqing and Alibaba Cloud, released "China's first large-scale autonomous driving scenario library based on vehicle-road collaborative cloud services". This is China's first use of real traffic data to generate and satisfy data The autonomous driving scene library that complies with compliance requirements will further accelerate the maturity improvement of China's autonomous driving and the collaborative development of vehicle, road and cloud.

Finally, MANA cognitive intelligence has ushered in a new stage. Through in-depth understanding of a large number of human driving across the country, learning common sense and personification of actions, the driver's assisted driving decision-making is more like actual human driving behavior. It can choose the optimal route based on the actual situation to ensure safety, and the body feels more like an experienced driver.

Based on insights into the era of autonomous driving 3.0 and the huge consumption of computing power required by large-scale model training, WeiMo also officially announced the WeiMo Supercomputing Center at the scene. "How to improve training efficiency, reduce training costs, and achieve low-carbon computing is a key threshold for autonomous driving to enter thousands of households." Gu Weihao said that the goal of the Haimo Supercomputing Center is to meet the requirements of large-scale models with hundreds of billions of parameters and training data scale With 1 million clips, the overall training cost is reduced by 200 times.

“We are passionate about innovation and open to new ideas, new methods, and new technologies. We pay special attention to technical routes that can form a positive cycle with the growth of data scale. This is also the first priority in making technical strategic decisions. Principle: A technical route that can quickly transform data scale advantages into capability advantages is a good route." Gu Weihao said that when it comes to the exploration and implementation of cutting-edge technologies, Wei Mo will always maintain the most active, keen and open attitude, and strive to Provide users with better product experience and promote the development and progress of the industry.

Haimo City NOH plans to mass-produce in September: Navigation-assisted driving that better understands China’s urban road conditions, and is confident that the user experience will be China’s first

in August 2022 At the Chengdu Auto Show, Weipai announced that the new Mocha DHT-PHEV lidar version equipped with Haimo City NOH is planned to be mass-produced in September and will be released within the year, and will be delivered as soon as it is launched.

Haimo City NOH is about to enter the city. What the industry is most concerned about is whether the autonomous driving function can cope with the complex traffic conditions in the city. "Haimo City NOH is a navigation-assisted driving that better understands China's urban road conditions." Gu Weihao said in his speech that Haimo City NOH adopts the "emphasis on perception, light on maps, and large computing power" technical route. MANA's six major milestones have been upgraded to With the smooth implementation of urban NOH, we are not at all confident that the user experience will reach a leading level. Gu Weihao highlighted five highlight functions at the scene: intelligent recognition of traffic lights, intelligent left and right turns, intelligent lane changes, intelligent obstacle avoidance - static, intelligent obstacle avoidance - dynamic. In addition, the "smart traffic flow processing" function It will also be officially released.

China's first 100,000-yuan terminal logistics automatic distribution vehicle "Little Magic Tuo 2.0" has been rolled out and delivered, and its large-scale implementation will speed up

As the "vanguard" in autonomous driving, terminal logistics Automatic logistics delivery vehicles are increasingly appearing on the streets and alleys to solve the "last mile" logistics and distribution problem. Driven by various factors such as policy, technology, and market, the industry is ushering in a large-scale explosion.

At the 5th HAOMO AI DAY held in April 2022, Feimo pioneered the industry and officially released China's first 100,000-yuan end-logistics automatic distribution vehicle product - "Feimo Little Magic Camel" 2.0”, the first shot in the scale battle for end-of-line logistics automatic distribution vehicles in 2022. Five months later, BoMo announced that "HaiMo Little Magic Camel 2.0" has been offline for delivery, further promoting the large-scale implementation of terminal logistics automatic distribution and accelerating the implementation.

According to Zhang Kai’s introduction in his speech, “Hao Mo Xiao Mo Tuo 2.0” has L4 level autonomous driving, remote driving, low-cost deployment, vehicle management platform, remote monitoring platform, order management platform, WeChat applet seven The large core functions are at the leading level in the industry, ensuring the low-cost deployment of Xiaomotuo 2.0 and the efficient execution of order distribution. In terms of commercialization, the bicycle is priced at 128,800, with an estimated annual output of 10,000 units, which can cover parks and urban open roads.

The most luxurious guest lineup in the history of AI DAY helps out, and Haomo’s technology and ecological influence rises again

"Haomo Moments" appeared to help out, and has always been the "favorite" of HAOMO AI DAY "Highlight Session", this event has the most luxurious guest lineup in the history of AI DAY: not only Zhang Yaqin, academician of the Chinese Academy of Engineering, professor of Tsinghua University, and dean of Tsinghua Intelligent Industry Research Institute (AIR), but also Vice President of Alibaba Group and Alibaba Cloud Computing platform leader Jia Yangqing, Cambrian CEO Chen Tianshi and other world's top AI scientists and entrepreneurs made appearances, as well as Weipai deputy general manager Qiao Xinyu, Qualcomm's senior director of product marketing Ai Hezhi; relevant government departments and professional institutions in Beijing, Shenzhen and other places Watch and communicate online and offline with partners such as Alibaba, Meituan, Tencent, NavInfo, AutoNavi, Sagitar, JD.com, BMW, Mercedes-Benz, and Haomo Investors.

The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs

(Hao Mo partners attended HAOMO AI DAY)

Zhang Yaqin shared his thoughts on autonomous driving in his speech. He said that autonomous driving will be a very challenging and complex task in the field of AI in the next five years, and it will also be an important technical force driving the transformation of the global automobile industry. "I think autonomous driving is achievable, but its long-term operation in an open environment still needs to withstand a series of technical and social acceptance challenges to solve safety and reliability issues. And on the path to implementation, vertical field autonomous driving will It will be launched faster and gradually spread to the general driving market."

Jia Yangqing, Vice President of Alibaba, shared his thoughts on AI engineering in the field of autonomous driving. He said that through the innovation and exploration of Alibaba and Haimo in the field of AI big computing and optimization, he saw huge potential for both parties in the future in the three fields of AI big computing, cloud collaboration and vehicle-road collaboration, as well as a broader exploration of the combination of data and AI. Opportunities for collaboration. In the future, we hope that Alibaba and Haomo will continue to strengthen close cooperation, continue to deepen their efforts in the field of autonomous driving, and jointly establish an autonomous driving intelligent system with industry-wide influence. Cambrian CEO Chen Tianshi said in the keynote speech "Car-Cloud Collaboration, Let Cars Better Understand and Serve Humans" that in the era of smart cars, whether it is the cloud or the car, the demand for computing power is rising exponentially, and computing power has become The core driving force for smart car upgrades.

Qiao Xinyu, deputy general manager of Wei Brand, said that Wei Brand is committed to creating the best solution for every scenario of smart driving in China. The arrival of Mocha City NOH intelligent assisted driving will promote the digital travel lifestyle and move from "high-speed scene intelligence" to the "urban scene intelligence" cycle.

Ai Hezhi, senior director of product marketing at Qualcomm, said that Qualcomm supports Haimo on the one hand to continue iterating on existing projects, and at the same time supports Haimo to continue to apply the results achieved with HPilot3.0 on Qualcomm’s new platform. One platform is expanded to more intelligent assisted driving platforms and applied to more models. Qualcomm will continue to support Wei Mo in realizing his broader ideals!

At the end of the event, Gu Weihao said: "More than 1,000 days ago, we were born in Haomo, and witnessed the fastest 1,000 days of autonomous driving in China. We are delighted with the achievements we have made. But 1,000 days , is just the beginning of the battle."

Zhang Kai said: "From faltering to smug, I never forget my original intention and mission. I don't stop when I reach the middle of the mountain, and I stride forward steadily on the new journey! In the past , now and in the future, all BiMo people will adhere to BiMo’s mission, values, and strategies, overcome all obstacles, and travel day and night to reach the starry sea of ​​autonomous driving.”

The above is the detailed content of The official announcement of the Haimo Supercomputing Center: a large model with 100 billion parameters, a data scale of 1 million clips, and a 200-fold reduction in training costs. For more information, please follow other related articles on the PHP Chinese website!

Statement:
This article is reproduced at:51cto.com. If there is any infringement, please contact admin@php.cn delete