In the second part of the science fiction novel "The Three-Body Problem", the author Liu Cixin describes the scene of the protagonist's thinking as follows: "He sits on a bench by the lake, looking at the reflection of the snowy mountains in the lake, thinking nothing or nothing, and before he knows it, the day has passed." ”
Later, the protagonist found a way to save the world, and this plot also became the inspiration for entrepreneur Gu Weihao. On December 23, 2021, as CEO of Milli Mobility, he released China's first autonomous driving data intelligence system MANA, Chinese named "Snow Lake".
Milli is running on a futuristic track.
Ideally, when a self-driving car is on the road, it is not a car, but a complex system of intelligent space and a set of rapidly generated, changing data interactions. This system is connected to the intelligent computing center with supercomputing capabilities through the mobile Internet, which can make instant judgments and control on the needs of passengers in the space or the actual situation of road traffic outside the space, and also has the ability to learn and evolve.
How high is the computing power requirement for the intelligent computing center?
At the beginning of the new year, new numbers try to give the answer, 6.7 billion, an 18-digit number - on January 5, the intelligent computing center "MANA OASIS" was officially established, which is used for autonomous driving large model training, with a computing power of 670PFLOPS (6.7 billion floating point operations per second), which is currently the largest intelligent computing center in China's autonomous driving industry.
This means a qualitative change, and helping to build an "oasis" in the snow lake is a new partner, Volcano Engine, which provides an intelligent computing cloud solution that increases the efficiency of model training by 100 times. At the same time, MANA OASIS has a storage bandwidth of 2 terabytes per second and a communication bandwidth of 800 gigabytes per second.
Cars have been driven in human history for more than 100 years, and their power comes from their engines. Today, the computing power "engine" will give the car the power to drive into the future.
"MANA OASIS gives Mumo supercomputing power, and with abundant data and computing power, the technical product capabilities of Momo will be more powerful." Gu Weihao said at the press conference, "The end will sprint into the era of autonomous driving 3.0. ”
Why 3.0?
At the 7th "Milli AI DAY" event held on January 5, Zhang Kai, Chairman of Milli End, made a prediction on the development of the autonomous driving industry in 2023.
In his opinion, this year, high-level intelligent driving will become the standard of mid-range models, a new generation of chips and sensors will enter the market, and the autonomous driving ecology will be richer and more diverse. At the technical level, the capabilities of large models in data synthesis and knowledge extraction will help improve the iteration speed of autonomous driving. When intelligent driving systems compete, functional competition will transform into commuting efficiency competition; Supercomputing centers will become the entry configuration for autonomous driving companies. At the product level, urban navigation assisted driving has entered the stage of heavy perception, the curtain of large-scale mass production and delivery has opened, and the user experience of intelligent driving will change from "early adopter" to "dependent".
The picture predicted by Zhang Kai shows that the 3.0 era of autonomous driving is fully opened, and the second half of the competition of related enterprises has also entered the acceleration stage.
On May 8, 2012, the Nevada Department of Motor Vehicles issued the first self-driving car license plate in the United States for Google's driverless car, with the license plate number "001" and the sign of "∞ (representing infinite mathematical symbols)", which symbolizes the future of automobiles. The car is modified from an ordinary car, and there is an additional laser rangefinder on the roof to determine the distance between the car and the surrounding objects. It is also equipped with 4 radars on the front and rear bumpers, cameras in the rearview mirrors, sensors on the left rear wheel, GPS and inertial test equipment on the body, which are used to collect information about the car's surroundings.
This is the 1.0 era of self-driving cars. Today, in Zhang Kai's view, the 1.0 era relies on millimeter-wave radar, cameras and the vehicle's own wire-controlled ability to realize functions such as single-lane autonomous cruising, which can be said to be the era of hardware driving. In the subsequent 2.0 era, without adding hardware, self-driving cars rely on software to strengthen the lateral control of the vehicle, which can achieve functions such as autonomous lane change and active dodge, which can be said to be the software-driven era.
What is 3.0? The logo is "data-driven."
Zhang Kai explained that its most typical feature is to let users really use data in the scenario to drive the iteration of autonomous driving. Just like any product, the core value of its existence is to serve the needs of people, and "people-centered" iterative evolution can stimulate the customer value of the vehicle intelligent driving system and ensure the accuracy of iteration.
"The better the product, the more users love to use", behind the simple truth, is the essential law and positive cycle.
For today's autonomous driving companies, the core resource is data, and the biggest problem is how to deal with the explosive growth of data efficiently and quickly. The vast array of product data, ultra-large-scale data training, and the application of large models are all challenging the computing power of processing data. To fight ahead in the industry, MANA OASIS undoubtedly has the technical strength to fight for the pinnacle.
No matter when you enter the track, you can be ahead
Momo Chi Heng is very young, and it was born for more than 1,000 days and nights. It is not the first company to rush into the self-driving track, but its "speed" is stable and fast.
So far, it has launched HPilot, a third-generation passenger car assisted driving product, and has undergone six OTA upgrades in two years. In terms of "perceptual intelligence", it "makes it possible for China to have no unrecognizable traffic lights and lane lines"; In terms of "cognitive intelligence", the original Snow Lake has replaced the traditional handwriting rules and parameters with machine learning models to make scene decisions more generalized and applicable. In November 2022, the first batch of 300 Mocha PHEVs equipped with HPilot set sail in Europe, becoming the first autonomous driving company in China to go overseas and mass-produced - China's mass-produced autonomous driving can withstand strict data security tests in Europe.
The "unicorn" has a unique "car company + technology" gene, firmly cultivates in the field of data acquisition and utilization, runs on the "gradual" route of automatic driving, and ranks among the first echelon.
Looking around, in the autonomous driving industry track, some people left the scene sadly, and there were also "luxury cars" rubbing shoulders. Technology giants such as Google Waymo, Baidu Apollo, Tesla, Xpeng and other car companies, as well as Pony.ai, Yushi Technology, etc. are racing against time. Living, accelerating, and leading are the goals of all these businesses. Some insiders said that "AI automatic driving is a protracted battle."
In 2022, the company will continue to rank first in China's mass production of autonomous driving, and the third-generation HPilot products will be equipped with nearly 20 models such as Wei brand, tank, Ora, and Great Wall Cannon, and the driving mileage of assisted driving users will exceed 25 million kilometers; Milli City NOH is China's first urban navigation assisted driving product that can be mass-produced, and the software cover has reached the delivery state; The terminal logistics automatic distribution vehicle has initially completed the commercial closed loop, delivering more than 1,000 units, and the number of delivery orders of Little Magic Camel has exceeded 130,000, and the commercialization process is accelerating in an all-round way; MANA has more than 420,000 hours of learning time, and the virtual driving experience is equivalent to 55,000 years of human drivers.
Data-driven technological progress, the harvest on this road, let the company shout "China's mass production autonomous driving industry first".
In October 2022, the international standard ISO 34501 "Vocabulary of test scenarios for autonomous driving systems of road vehicles" was officially released. The Ministry of Industry and Information Technology stated that it will organize relevant units to take the principle of synchronous research and formulation of international and domestic standards, give full play to the advantages of rich application scenarios and active technological innovation of the mainland automotive industry, accelerate the research and development of advanced autonomous driving technology, the construction and application of basic scene libraries, and the capacity building of industry testing and evaluation, and continue to improve the mainland's participation and contribution in the construction of international standards and legal systems and industrial integration development. At the same time, for the first time, the mainland issued a map license for advanced driving assistance in cities.
The time of opportunity has come. The industry generally believes that 2020-2030 will be the "golden decade" of autonomous driving development, and according to McKinsey, China is likely to become the world's largest autonomous driving market in the future, and it is expected that by 2030, China's autonomous driving-related new car sales and travel services will generate more than 3.58 trillion yuan.
In the face of the huge market and the struggle in the industry, Momo Chi Heng welcomed a partner.
Put on a "volcano"-like engine
This is a crazy speed-up, and it is also an inevitable choice.
In fact, any autonomous driving company that has developed to a certain scale and accumulated data to a certain volume requires an intelligent computing center to process data and solve problems such as model training.
In 2021, Tesla showed the world the supercomputing center Dojo; In 2022, Xpeng Motors cooperated with Alibaba Cloud to create the "Fushake" Intelligent Computing Center in Ulanqab, Inner Mongolia.
At the beginning of 2023, the addition of Volcano Engine helped the "Snow Lake Oasis" to become the largest autonomous driving intelligent computing center in China. When introducing this new "engine", CEO Gu Weihao once said that just as soldiers need weapons to go to the battlefield, the weapons of AI engineers are in the cloud, one is data, the other is cloud resources. Compared with the "public cloud", the intelligent computing center is more targeted and cost-effective, so that the cloud resources are more in place and the combat effectiveness is more sufficient.
The changes brought by the new "engine" to the company are obvious. The first phase of the intelligent computing center delivered more than 2,000 GPU cards, and this number continues to grow; Based on Sparse MoE, 100 billion parameter large model training can be easily completed through cross-machine sharing, and the training cost of one million clips is only 100 calories per week, reducing the training cost by 100 times.
The "five models" that Milli Zhixing is proud of have also been assisted by the Intelligent Computing Center, comprehensively upgraded, and the iteration of mass production of autonomous driving products has accelerated, and it is expected that in the first half of 2024, the NOH landing of Milli City will reach 100 cities, and fully unmanned driving will be achieved in 2025.
NOH is China's first mass-produced urban navigation assisted driving product. Compared with other competing products, the biggest feature of Momo is "heavy perception, light map", that is, navigation assistance driving without relying on high-precision maps.
The high-precision map, which is known as the "opening of the eyes" of the intelligent driving system, is of great help to route decision-making, but it is easily limited by cost and policy, and large-scale promotion is challenging. The urban NOH mainly relies on the perception route of cameras and lidar, which needs to collect a large amount of data, and then operate, analyze, and decode through the intelligent computing center to improve the technical capabilities of the system.
The addition of the volcano engine optimizes the computing power of "Snow Lake Oasis" to the extreme. "With the help of five models, MANA's latest vehicle-end perception architecture is integrated from multiple downstream tasks scattered in the past to form a more end-to-end architecture, including general obstacle recognition, local road network, behavior prediction and other tasks, and the vehicle-end perception architecture has achieved cross-generational upgrades." Gu Weihao said. This also means that the perception ability of the milli is stronger, the product power is stronger, and the acceleration towards full unmanned driving.
In the future, Volcano Engine will give full play to its advantages in cloud services, AI, data and other technologies to help the sustainable growth of its business, jointly promote the leapfrog development of intelligent training platforms in the autonomous driving industry, and accelerate the agile iteration and commercialization of autonomous driving technology.
As a cloud service platform owned by ByteDance, Volcano Engine has a leading advantage in data-driven. Based on the large computing power scenarios of car companies and the current problems encountered by car companies in external computing power, Volcano Engine provides end-to-end intelligent computing cloud solutions for car companies.
In 2023, the "four major battles" will be launched, including "the battle of the king of intelligent driving installed capacity", "the battle of the peak of the MANA big model", "the battle of 100 cities in the city" and the "battle of the terminal logistics automatic distribution business" - young express car, the new engine has been launched, and the industry will wait and see.
It is reported that "Snow Lake Oasis" is located in Datong, Shanxi, a land that gave birth to ancient human civilization and carries a long history and culture, and also nurtures the future.
#自动驾驶最强算力突破67亿亿大关 #