At the beginning of 2023, the first is the beginning of the year, milli-milli Zhixing held HAOMOAIDAY, put out the automatic driving industry's largest intelligent computing center, and then there is a small Peng, the ideal new year full letter of the sword refers to the city navigation assisted driving, followed by dialogue AI big model ChatGPT fire all over the network, automatic driving AI technology has once again become the top stream.
Whether it is the "city" of automatic driving, or the "evolution" of ChatGPT, behind it all is the exponential growth of data and arithmetic demand, as well as the training of large models. When the demand comes up, the Smart Computing Center, as the "new infrastructure" for autonomous driving, is mentioned more and more by the industry.
The Smart Computing Center is an intelligent computing center that is based on the theory of artificial intelligence and adopts the leading AI computing architecture to provide a new type of public **** computing infrastructure for computing services, data services and algorithmic services required for AI applications, in other words, the Smart Computing Center is in fact a computing power supply and production platform. The company's main goal is to provide the best possible service to its customers," he said.
Subtracting the marginal cost of automated driving to the cloud
Some people say that the smart computing center is an enabler for the development of automated driving because automated driving is the most important part of the process, and it's the most important part of the process. The training of driving algorithm models is one of the typical scenarios of machine learning, and its visual detection, trajectory prediction and traffic planning algorithm models need to complete highly concurrent parallel computation at the same time, which has a very high demand for computing power, and Smart Computing provides huge computing power to improve the maturity of algorithm models.
In the field of automatic driving, when it comes to the Smart Computing Center, Tesla has to be mentioned first.In 2017, after the appearance of the Transformer network, it laid the foundation of the current mainstream algorithmic architecture in the field of large models, and then, in 2020, Tesla introduced the Transformer large model into the field of automatic driving, which is the beginning of the application of the AI large model to automatic driving. After this, Tesla began to work on building its own AI computing center, Dojo, which used a total of 14,000 NVIDIA GPUs to train AI models. In order to further improve efficiency, Tesla released its own AI acceleration chip, D1, in 2021, and plans to package 25 D1s together to form a training module (Training tile), which will then form a cabinet (Dojo ExaPOD). In a recent issue of Tesla AI DAY, Musk said that the Tesla supercomputer cluster ExaPOD will be deployed in the first quarter of 2023.
Domestically, in August 2022, Xiaopeng Automobile and Aliyun jointly built the largest domestic automatic driving intelligent computing center "Fushang", which is specifically used for automatic driving model training. Automatic driving model training, the arithmetic scale of 600PFLOPS, equivalent to 6 billion floating point operations per second can be completed. However, this record has only lasted for more than four months.
January of this year, the milli-milliard Zhixing joint volcano engine, *** with the launch of the automatic driving industry's largest intelligent computing center MANA OASIS (snow lake - oasis), 6.7 billion floating-point operations per second, storage bandwidth of 2T per second, communication bandwidth of 800G per second. geely also in January 28 on-line geely starry intelligent computing center, has access to the intelligent driving and car networking experiment data Nearly one hundred PB, concurrent computing support for online vehicles up to one million.
From the point of view of the existing situation, the two factors of cost and demand are the attractive features of the Smart Computing Center.
At the cost level, arithmetic power, as a basic element of autonomous driving, requires higher performance of smart computing centers for training, labeling and other tasks. Taking the MANA OASIS of MilliMax as an example, by deploying Lego high-performance arithmetic libraries, ByteCCL communication optimization capabilities, and large model training frameworks, software and hardware are integrated, and MilliMax optimizes the arithmetic power to the extreme. In terms of training efficiency, based on Sparse MoE, through the cross-machine **** enjoyment, easy to complete hundreds of billions of parameters of the large model training, and millions of clips (millimeters of the smallest video annotation unit) training cost is only a hundred card week level, training costs reduced by 100 times.
Building an efficient, low-cost data intelligence system is the foundation for the healthy development of automated driving technology, and an important part of the automated driving system that can continue to iterate and move forward, and even more importantly, the key to the closed loop of automated driving commercialization.
He Xiaopeng, chairman of Xiaopeng Automobile, has stated, "If we don't reserve arithmetic power in advance in this way now (Smart Arithmetic Center), then in the next five years, the cost of enterprise arithmetic power will be added from the billion level, to the billions of levels."
If you continue to use public cloud services, the rising marginal cost is only one aspect, but more importantly, the Smart Computing Center allows self-driving companies to realize the "dedicated cloud". The development of automatic driving includes from data collection to data screening, marking, model training, playback verification, simulation testing and so on. The essence of cloud computing is to rent computing equipment, cloud service providers are unified procurement of equipment, in order to obtain more customers, these devices have a great deal of commonality, equipment used within the CPU, GPU / AI gas pedal, memory models and specifications are relatively fixed, it is difficult to form an optimal match with the algorithms of the car companies and automated driving companies. Moreover, cloud service vendors do not have a high level of understanding of autonomous driving algorithms, which inevitably leads to loss and inefficiency in scheduling arithmetic. Therefore, from the perspective of demand, it seems that the Smart Computing Center can be a bottoming-out godsend for autonomous driving and car companies.
Likewise, with the support of MANA OASIS, the new debut and upgrade of the five models of MANA, and the cross-generation upgrade of the vehicle-end perception architecture, the layout of the technology stack of millimeters continues to maintain a complete leading posture, especially in perception, cognition and other levels of the industry, leading the direction of development of the big model, big arithmetic, big data, and sprint into the era of automatic driving 3.0.
Taking data collection, screening and labeling as an example, the automatic driving system needs to collect a large amount of road environment data in the pre-development stage so that the vehicle can quickly and accurately identify key information in the driving environment such as lanes, pedestrians, and obstacles just like a human driver. The only way to do this is through repeated training and validation on the basis of massive data, the vehicle's cognitive level of the road environment gradually converges to the real scenario, and the accuracy of judgment is continuously improved in the process.
Not only that, the data collected by the car companies also need to be trained for the model, the algorithm generates the model by operating on the data, and the Smart Computing Center will be the gas pedal that drives the training of the big model and the massive data. Based on Sparse MoE, MilliMax carries out sparse activation according to the calculation characteristics, improves the calculation efficiency, realizes the effect that a single machine with 8 cards can train a large model with 10 billion parameters, realizes the method of cross-machine*** enjoying exper, completes the training of a large model of 100 billion parameter scale, and reduces the training cost to the level of a hundred-card-week; MilliMax designs and realizes the industry-leading multitasking parallel training system, which is capable of simultaneously processing The industry-leading multi-task parallel training system can simultaneously process information from multiple modalities such as images, point clouds, structured text, etc., which not only ensures the sparsity of the model, but also improves the computational efficiency; MANA OASIS training efficiency has been improved by 100 times.
Gu Weihao, CEO of Milliwheel, also explained the underlying logic of building the ICC in detail: "The first requirement of automatic driving for the ICC is definitely computing power. The mega computing power of the Smart Computing Center represents how many AI engineers are able to make what big models and how many big models can be trained in this practice arena."
Intelligent assisted driving "into the city" MANA OASIS to help millimeters to solve what difficulties?
Now many car companies and self-driving technology companies have begun to make the creation of smart computing centers the next stage of competition. At the HAOMO AI DAY in January this year, Zhang Kai, chairman of millimetre wise, gave ten new predictions for the trend of the automatic driving industry in 2023, and supercomputing center was listed among them, "Supercomputing center will become the entry configuration of automatic driving enterprises."
In fact, nowadays, as new energy vehicle brands have generally listed assisted driving in highway scenarios as standard, the playing field has quietly shifted from the highway to the city. Compared with high-speed navigation assisted driving, urban driving involves a series of difficulties such as traffic lights, intersections, pedestrians and electric vehicles, obstructions, fixed obstacles, frequent braking and starting, and so on, and the complexity has been raised by several orders of magnitude.
If only the actual test vehicle to challenge these urban scenarios can not be exhausted Corner Case, the cost, security, time will become a barrier to the development of enterprises. As a result, virtual simulation becomes the key to solving part of the cost and scene diversity, in which large-scale long-tailed scenes require data centers to provide sufficient arithmetic support. At the same time, the simulation scene to the reality of the return process, also need to provide support for the huge arithmetic power.
With the support of MANA OASIS, the five models of MANA, the data intelligence system of millimeters, are newly unveiled and upgraded. And with the help of the five models, MANA's latest vehicle-side perception architecture, from the past dispersed multiple downstream tasks integrated together to form a more end-to-end architecture, including general obstacle identification, local road network, behavior prediction and other tasks, millimax vehicle-side perception architecture has achieved cross-generation upgrade. This also means that millimeters have stronger perception capabilities and stronger products, accelerating towards full driverlessness.
Firstly, the visual self-supervised large model allows MilliMax to be the first to realize automatic labeling of 4D Clip in China. Using massive videoclip, MilliMax pre-trains a large model by video self-supervision, and uses a small amount of manually labeled clip data for Finetune (fine-tuning) to train the detection tracking model, making the model have the ability to automatically label; then, the original video corresponding to the ten million single-frame data that has already been labeled is extracted and organized into clips, in which 10% are labeled frames and 90% are unlabeled frames, and then input these clips into the model to complete the automatic labeling of 90% unlabeled frames, thus realizing 100% automatic transformation of all single-frame labeling to clip labeling, and at the same time, reducing 98% of clip labeling cost. The generalizability of the millimetre-end video self-supervised large model is so effective that it can accurately complete the automatic annotation even in some very difficult scenarios, such as severely occluded cyclists, small targets in the distance, bad weather and lighting.
Secondly, the 3D reconstruction of large models to help millimeters to do data generation, with lower costs to solve the data distribution problem, to improve the perception of the effect. Facing the industry problem of "accumulating corner cases from real data is difficult and expensive", MilliMax applies NeRF technology in the reconstruction of auto driving scene and data generation, which generates high realism data by changing the view angle, lighting, texture material, and realizes the low-cost acquisition of normal cases, and generates various high-cost corner cases. The data generated by 3D reconstruction of large models is not only more effective than the traditional manual explicit modeling and then rendering texture methods, but also less costly. The addition of NeRF-generated data also reduces the perceived error rate by more than 30%, and data generation can be fully automated without any human involvement.
Multimodal mutually supervised large models, on the other hand, can accomplish generalized obstacle recognition. After successfully realizing the accurate detection of lane lines and common obstacles, Milliwheel is thinking about and exploring more general solutions for the stable detection of many kinds of shaped obstacles in the city. Currently, millimax's multimodal mutually supervised large model, which introduces LiDAR as a visual supervisory signal, directly uses video data to reason about the generic structure representation of the scene. The detection of this generic structure can well complement the existing semantic obstacle detection and effectively improve the pass rate of the autonomous driving system in complex urban conditions.
The dynamic environment macromodel, which can accurately predict the topological relationship of the road, allows the vehicle to always drive in the correct lane. Under the route of heavy perception technology, HILLMOTH is facing the challenge of "real-time deduction of road topology" in order to minimize the dependence on high-precision maps. In order to minimize the dependence on high-precision maps, HMD faces the challenge of "real-time deduction of road topology". Based on the feature map of BEV, HMD takes the standardized map as the guidance information, and uses the autoregressive coding and decoding network to decode the features of BEV into the sequence of structured topology points, so as to realize the prediction of lane topology. It allows the perceptual ability of millimeters to realize the real-time inference of road topology just like human beings with the navigation hints from standard maps.
Millimeter believes that solving the intersection problem actually solves most of the urban NOH problems. Currently in Baoding and Beijing, the accuracy of topology inference of 85% of intersections is as high as 95%. Even for very complex and irregular intersections, millimeters can accurately predict them, more than an old driver.
The human-driven self-supervised cognitive grand model has been officially upgraded to DriveGPT in February this year, which is also the world's first automatic driving cognitive grand model. It can make millimeters driving strategy more anthropomorphic, safe & smooth. At present, MilliMax DriveGPT has completed the model construction and the first phase of data run-through, the parameter scale can be compared to the level of GPT-2. Next, DriveGPT will continue to introduce large-scale real takeover data, through the reinforcement learning of human-driven data feedback, to continuously improve the evaluation effect, and also use DriveGPT as a cloud-based evaluation model to assess the driving effect of small models on the vehicle side.
Simulation testing can effectively shorten the technology and product development cycle and reduce R&D costs. Typical long-tail scenario problems in the industry are not abundant enough, and extreme scenarios that are unattainable in reality can be conveniently generated using the simulation platform. As the simulation environment in the simulation test needs to realize multi-modal fusion to support the complexity of the sensor module, and thus also requires the support of large computing power.
In addition to millimeters, Tesla's supercomputing center, with nearly 20,000 GPUs, has had an immediate effect on the efficiency of autopilot training, maximizing the development efficiency of autopilot systems; Continental's high-computing power clusters, which have shortened the development cycle from weeks to hours, have enabled the implementation of autopilot in short- and medium-term business plans; the shortening of machine-learning time has accelerated the pace of new technology The shortening of machine learning time accelerates the speed of new technology into the market; "Fushang" supports the training time of Xiaopeng's core model of automatic driving from 7 days to within 1 hour, dramatically speeding up by nearly 170 times ......
Currently, an indisputable fact is that car companies with long-term plans in the field of automatic driving, whether they are making cars or cars for the future, have a lot to offer to the market. Vehicle enterprises with long-term plans in the field of automated driving, whether they are new car-making forces or traditional brands, or technology suppliers, are building their own supercomputing centers in order to grasp stable computing resources, shorten the development cycle, and accelerate the launch of automated driving products. On the contrary, if there is no supercomputing center, then the speed of automatic driving training will be significantly slowed down, and the gap between automatic driving companies will become more and more obvious.
Creating a data moat with an intelligent computing center The new digital infrastructure is gradually becoming a development "standard"
Automatic driving has developed so far, and the industry has found that intelligent assisted driving in passenger cars is the most likely business scenario to be spread on a large scale. According to data from the GIGA Intelligent Vehicle Research Institute, in 2022, the passenger car front-loaded standard equipment equipped with L2-level assisted driving rate in the Chinese market (excluding import and export) has exceeded 30% for the second consecutive month. The global penetration rate of L2 autonomous driving in new cars is expected to reach 53.99 percent by 2025, according to data from Wisdom Research Consulting.
This year, city navigation assisted driving has also started the journey of mass production. Western Securities predicted that from 2023 to 2025, the domestic market equipped with city navigation assisted driving models will reach 700,000, 1.69 million and 3.48 million vehicles, accounting for 17%, 40% and 70%, respectively.
In the context of accelerating the landing of urban navigation-assisted driving, the program of heavy perception, which is easier to copy and expand, has received more attention. Under the heavy perception technology route, in the face of the challenge of "real-time deduction of road topology", the choice of the millimeters is based on the feature map, the refined map as the guidance information, the use of autoregressive coding and decoding network, through the structured topology of the sequence of decoding points, to achieve the topological prediction of the lanes. From this, it is easy to see that the industry is gradually reaching *** knowledge of the heavy perception route, compared with the high-precision map program, more dependent on the arithmetic power to add.
Artificial intelligence is a gas pedal of innovation, and the Smart Computing Center can provide support for all kinds of technological innovation. On the one hand, ICC can provide arithmetic facilities to support the construction of a safe, trustworthy and reusable technology research and development environment, and provide intelligent computing services to accelerate the process of technology research and development for various fields of science and technology research and development; on the other hand, ICC is an integrated application carrier of new-generation information technology, and the rapid construction, promotion and large-scale application of the ICC will promote the rapid iteration of technology such as communication service networks, big data and artificial intelligence, thereby promoting technological innovation. On the other hand, ICC is an integrated application carrier of new-generation information technology. Autopilot data is fragmented, characterized by many small files, reaching tens of billions, and the training needs to exchange more data, the Smart Computing Center can provide sufficient bandwidth, and can allow the autopilot model to have a better parallel computing framework, and make use of all the hardware resources during the training.
On April 20, 2020, the National Development and Reform Commission (NDRC) for the first time clarified the scope of new types of infrastructure, which includes arithmetic infrastructure represented by smart computing centers.On January 10, 2023, the National Industrial Information Security Development and Research Center (NISDRC) launched the "Prospect Report of Smart Computing Centers in the Era of 2.0," which pointed out that after more than five years of development, smart computing centers are moving from the 1.0 rough expansion stage to 2.0 fine planning stage.
According to relevant statistics and calculations, more than 30 cities are currently building or planning to build smart computing centers, and the compound annual growth rate of the scale of China's smart computing power in the next five years will reach 52.3%. The innovative development of ICC will further consolidate the "arithmetic base" for AI and become a new engine to drive the rapid development of AI and related industries.
"We calculate that the cost optimization brought by the ICC is amazing, and will reach the level of hundreds of millions of yuan." This is the prediction made by Zhang Kai in January this year. From the current and future planning of mass production scale, millimeters of self-built IQC can save huge costs; at the same time, the efficiency improvement it brings is also very obvious.
Artificial Intelligence is developing rapidly, new algorithms are emerging, new technologies and models need to be introduced as soon as possible, at the same time, data is the biggest driving force for the development of intelligence, but also occupies a large number of cost components. With self-built ICC to create a data moat, not only can improve the industrial intelligence ecosystem, but also allow enterprises to occupy a first-mover advantage in intelligence, ICC as a new digital infrastructure, the future will inevitably lead the continuous iteration and upgrading of automated driving technology.
This article is from the author of the car industry vision, copyright to the author, any form of reproduced please contact the author.