The second half of autonomous driving: the kingdom of computing power, the big model of the car

author:Wisdom Valley Trends

Wisdom Valley Trend (ID: zgtrend) | Wang Yanhe


$500 billion shock

Storms are always sudden.

On September 11, Tesla's stock price rose 10%, and its market value increased by $80 billion, or about 580 billion yuan, overnight.

It only makes traders frightened.

The reason for the sudden rise was not Musk's new 370,000-word biography, but a 66-page report from Morgan Stanley.

Wall Street analysts are eyeing neither the new Model 3 nor the Mexican Gigafactory, but the Dojo supercomputer.

Afraid that others will not understand, Morgan Stanley drew a pie: Dojo is valued at more than $500 billion!

What is Dojo? AI computing power infrastructure for Tesla's autonomous driving.

To develop autonomous driving, it is necessary to collect billions of kilometers of driving data and train optimization algorithms.

With each level, such as L4 and L5, the demand for data and computing power for autonomous driving will increase by orders of magnitude.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

In order to improve training efficiency, Tesla developed its own supercomputer, which was officially announced at AI Day in 2021, called Dojo. Even in order not to rely on NVIDIA, he also developed his own D1 chip.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

Dojo's computing power reaches 1 EFlops, which is 10 billion billion floating-point operations per second.

Tesla will also spend $1 billion to expand to 100 EFlops in October 2024, equivalent to 300,000 Nvidia's A100 chip computing scale.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

People can't help but think of the world's largest cloud service provider, Amazon Web AWS.

At the beginning, Amazon reserved a large number of redundant servers for the e-commerce business and the Black Friday Shopping Festival, and the usually idle computing power was packaged and sold.

Today, AWS Cloud contributes 70% of Amazon's net profit, far surpassing e-commerce retail and making it its most profitable business.

Is it possible that Dojo is Tesla's AWS?

If Tesla also exports computing power, automatic driving technology, visual algorithms, and AI capabilities, it will harvest an ecology that is more valuable than cars.

So, Morgan Stanley wrote:

Dojo gives Tesla an "asymmetric advantage" in the $10 trillion field of autonomous driving.

Obviously, not all car manufacturers, all Tesla.

American auto giants such as Ford and General Motors suffered a strike of 13,000 auto workers. Workers demanded a raise in fear, fearing "harvested by new technology." Because the automated production of electric vehicles will make a large number of traditional workers unemployed.

Global automakers are also caught in the bottleneck of slow transformation, price reduction competition, and profit reduction.

In other words, few people can take hundreds of millions of dollars in cash flow to build computing power infrastructure, regardless of cost.

Huawei has estimated that L4 autonomous driving requires at least 1 billion kilometers of road testing. That's equivalent to 1 million cars, running 10 hours a day for 1 year in a row.

Xiaopeng once revealed that the annual computing power expenditure exceeds 1 billion yuan; In the past two or three years, the demand for computing power of domestic autonomous driving companies has also doubled.

Among car manufacturers, a consensus is being formed: instead of doing it yourself, it is better to cooperate deeply with cloud service providers.

Why not huddle and become a more deterministic community?


Four wheels + a supercomputer

The matter of building a car used to be called four wheels plus a sofa;

Now it's four wheels, plus a supercomputer.

As the next generation of mobile terminals, smart cars are all on the cloud, which is a repeat of the opportunity of "smart phone + mobile Internet".

Tesla wants to rule the era of autonomous driving, and the opponent it can fight can only be born in China.

The electrification penetration rate in the Chinese market is the leading in the world;

In the first half of 2023, China's automobile exports surpassed Japan, becoming the world's largest automobile exporter.

According to the China Association of Automobile Manufacturers, Continental's automobile exports from January to July reached 2.533 million units, a year-on-year increase of 67.9%, a record high. It is expected that the export of new energy vehicles will exceed 1 million units throughout the year.

Large domestic cloud service providers such as HUAWEI CLOUD, Alibaba Cloud, and Baidu Cloud have also launched automotive cloud-related services.

With a larger market, a closely coordinated industrial chain, and the participation of many developers, China has every reason to be born: ultra-large-scale automotive cloud, top-notch autonomous driving technology.

For example, Huawei has established a cloud data center in Ulanqab and opened a special automotive zone to provide a cloud infrastructure with security, compliance, independent innovation, and surging computing power for autonomous driving scenarios.

HUAWEI CLOUD Ascend AI Cloud Service is deployed in the automotive zone, which provides 30 days of uninterrupted kcal training and no more than 10 minutes of breakpoint recovery.

This is the mainstream mode of cooperation between car companies and cloud service providers.

Car companies are like dispatching water and electricity, obtaining computing power resources. Cloud service providers solve the efficiency problems of computing power, storage, and network transmission.

However, the cloud requirements of automotive companies will also be upgraded from simple computing power infrastructure to software services and platform ecosystems.

HUAWEI CLOUD has been working with automotive companies for a long time to build an autonomous driving R&D platform, which can provide both "one-stop" and "building block" deployment solutions.

For start-up automakers, HUAWEI CLOUD provides a one-stop data closed-loop and autonomous driving expert team, enabling automakers to quickly build a complete autonomous driving R&D platform from 0 to 1.

Of course, each car company has different development progress, and the cooperation needs are naturally different.

For example, some car companies want a "sense of boundaries" and are unwilling to deeply bind with cloud vendors; Car companies with their own data centers need more software cooperation.

In this case, the "building block" is more flexible, and HUAWEI CLOUD provides an open source platform, integrated reference code, and a variety of toolboxes, which car companies can choose according to their needs.

"The key factor affecting the landing of high-level autonomous driving is no longer to solve common general cases, but to allow models to quickly learn various uncommon but constantly emerging long-tail problems." Dr. Jin, an AI expert at HUAWEI CLOUD, explains.

This is the most headache for self-driving companies.

At present, autonomous driving technology can cope with 95% of driving scenarios, but the remaining 5% of edge scenarios, or corner cases, are scenarios that have never been learned.

Identifying a new Corner case takes weeks or even months to collect tens of thousands of samples.

In theory, to achieve complete autonomous driving, at least 10 billion kilometers of road test data must be accumulated. Cost and time are unacceptable.

To solve this problem, HUAWEI CLOUD's Pangu car model can generate a virtual space based on multi-journey data collection.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

Parameters such as objects, positions, spatial layouts, and movement trajectories of traffic participants in the virtual space can be adjusted.

This solution can shorten the time of Corner case data acquisition and training from weeks to two days, which greatly improves efficiency.

This is also the first time that autonomous driving has used the ability to generate data, and the capabilities of Pangu's large model are not only there.

Often, cloud providers are new to the industrial manufacturing space.

However, Huawei itself is a representative of advanced manufacturing, and the Pangu Auto model can cover all scenarios such as automobile design, manufacturing, and marketing, so that each employee has an AI expert assistant.

For example, FAW Jiefang has been deeply integrated into the Pangu model.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

Previously, they designed the appearance of car seats, which took two weeks to make renderings, but now they use large models that can be generated in seconds, quickly iterating to satisfaction, and achieving "what is said is what you get".

For example, there are more than 80,000 standards for car design, 1.6 million pages of instructions, and novice designers spend weeks combing through the standards, while large models can quickly locate relevant chapters and give the source of standards.

After the Pangu model is integrated into FAW Jiefang's data, it can provide all-round empowerment for Jiefang, covering R&D, production, supply and marketing services.

The same is deep in automatic driving, AI computing power, Huawei and Tesla are different.

As China's top technology supplier, Huawei has more to shoulder.

The car model is a key corner of the chessboard, but you have to see the direction of the whole chessboard:

An industry that is caught in computing power anxiety, and not only car companies?

Thousands of industries, 40 million companies, who does not need intelligent transformation and find the certainty of growth?


Reopening the world

After 2 months, ChatGPT turned one year old.

As of July, China has a total of 130 big models out, who can go further?

From the 100-model war to the bubble reveal, the new consensus is: the focus of the big model is on the industrial landing.

To study artificial intelligence, algorithms, computing power, and data are just needed, but each is a money-burning product.

Not all tracks are like autonomous driving, and there are car companies and consumers who pay.

Investors tighten their wallets, advocate cash as king, and are especially cautious in their deals.

Even OpenAI is showing signs of fatigue, launching an enterprise version of ChatGPT to compete for the B-end market, hoping to make profits faster.

The new world came and was pulled back to reality by financial indicators, technical difficulties, and cost control.

As the filter of "ChatGPT moments" fades, companies also see the limitations of large models.

For example, the accuracy rate of large models is not enough, 80% may meet daily use, but it is not enough for enterprises, especially in industries such as law, medicine, finance, etc., which have almost zero tolerance for errors.

For example, the cost of large model deployment and technical threshold are prohibitive. What the enterprise side wants is always low-cost, ready-to-use, one-stop service products.

Thousands of industries clearly have real needs and real problems;

AI companies also want to land in the industry to obtain data flow and cash flow;

But between the two, there is a gap - the gap between technology and industry.

Huawei has always believed that solving problems, doing difficult things, and taking root in the industry.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

For example, this year, the whole country from the northeast to Beijing-Tianjin-Hebei, from Guangdong, Hong Kong and Macao to eastern China, has suffered heavy rainfall.

Headlines such as "Rainstorm of the century", "Once-in-a-century", "One day and a year of rain" flooded social media, affecting millions of people and causing countless economic losses.

Some people question, is our weather forecasting system useful? Why not give advance warning and allow time for disaster prevention?

This is the gap between technology and reality.

Modern weather forecasting technology, in the simplest words: collect a large amount of meteorological data, input it into a supercomputer, and simulate the prediction results through complex algorithms.

Sounds like this is an area where AI excels? That's right.

This summer, the HUAWEI CLOUD team worked around the clock to upgrade the Pangu Meteorological Model.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

Pangu predicts the path of Typhoon Dusuri

It not only performs accurately in the prediction of this year's typhoon path, but also realizes the precipitation forecast for the next 6 hours and 24 hours.

The Pangea Meteorological Model has become the first AI model with accuracy exceeding traditional numerical prediction methods. At present, cooperation has been established with the National Meteorological Administration of China, Shenzhen Meteorological Bureau, European Medium-Range Weather Forecasting Center and Thailand Meteorological Bureau.

With just one server, 10 seconds, it can predict the global typhoon path in the next 10 days;

Next, Pangu wants to challenge the red warning of heavy rain, from 3 hours in advance to 24 hours in advance.

This is a lifeline to seize time, using human technology to confront the chaos and disorder of nature.

On September 20, Huawei CONNECT 2023 was held, with the theme of "Accelerating Industry Intelligence" and working with enterprises in thousands of industries to move towards an intelligent world. At the conference, the latest progress of the Pangu Grand Model was also revealed.

The second half of autonomous driving: the kingdom of computing power, the big model of the car

At the model level, the Pangu model forms a three-layer architecture of 5+N+X.

The L0 layer at the bottom is 5 base large models; The L1 layer is a large model of N industries such as automobile, mining, meteorology, drug molecules, government affairs, and digital humans; The L2 layer is a scenario model for X specific businesses, such as supply chain logistics and typhoon paths.

For example, it takes 10 years and costs $1 billion to develop a new drug, which is the famous "Double Ten Law" in the pharmaceutical industry.

Through the Pangu drug molecule model, the lead drug development cycle can be shortened to 1 month, and the R&D cost can be reduced by 70%.

Professor Liu Bing of the First Affiliated Hospital of Xi'an Jiaotong University developed a super antibacterial drug with the assistance of the Pangu drug molecular model.

In the past few years, Huawei has sent groups of doctors, experts, and scientists to go down mines, factories, live broadcast rooms, and meteorological bureaus.

Only when you get to the front line do you know what all walks of life need. Innovation is not about working behind closed doors, but about moving forward in solving problems.

We also know that everyone is more concerned about the problem of computing power.

After all, NVIDIA's A100/H100 has been cut off from domestic supply, and even the "castrated version" of A800/H800 has been speculated to sky-high prices, and it is difficult to find one.

In the next 10 years, the demand for AI computing power may increase by 500 times, can domestic AI computing power withstand the gap?

Huawei has long been prepared for this.

They have established cloud data centers in Ulanqab, Gui'an, and Wuhu to support the computing power of ultra-large-scale clusters.

In addition, from hardware to software, from computing power, operator libraries, AI frameworks, and AI platforms, Huawei has implemented full-stack self-research. It supports customers to smoothly migrate from other platforms to the HUAWEI CLOUD Ascend AI cloud service ecosystem.

Simply put, HUAWEI CLOUD's localized AI computing power allows everyone to not worry about computing power being "stuck in the neck" again.

With the guarantee of independent technology and the ecosystem rooted in the industry, HUAWEI CLOUD's large-model industrial revolution can be regarded as a firm foothold and a new game.



《奥本海默》里说,It's not a new weapon,it’s a new world.

Perhaps in the face of the big model, we should also say, It's not a new tech, it's a new world.

This rushing new world, whether you call it the fourth industrial revolution or the intelligent revolution, everything will be changed by AI, as if reopening the world.

The technology we seek must be disruptive innovation that breaks the original pattern and brings new growth certainty; It must help ordinary people cross the chasm, not deepen barriers.

Companies that can bridge the chasm, give birth; Enterprises that can benefit thousands of industries, strong.

The gears of fate have turned with a bang.

Read on