Holding the "Oasis" of the intelligent computing center in hand, Hao Mo is on the front line of the long-distance autonomous driving race between China and the United States.

What does the Intelligent Computing Center do?

By 2022, data-driven development will become the consensus for autonomous driving evolution. Along the data-driven route, autonomous driving is accelerating into the era of intelligent computing.

The Intelligent Computing Center came into being.

In fact, Tesla was the first to introduce intelligent computing into autonomous driving. It was the first to release Dojo, a supercomputing center dedicated to autonomous driving training. Since then, a number of players such as Haomo and Xiaopeng have announced the establishment of intelligent computing centers. .

What are the major intelligent computing centers competing against?

The answer is how to achieve closed-loop driving of massive data in a low-cost and efficient way.

"Large models of autonomous driving accelerate and reduce development costs, and the data is closer to the human driving environment. Simulation technology based on mass-produced real vehicle data will be combined with the ultra-large model of autonomous driving in the cloud to usher in new breakthroughs. The supercomputing center has become the entry-level configuration for autonomous driving. "Zhang Kai, chairman of Haomo Zhixing, said in the top ten autonomous driving trend predictions for 2023.

On January 5, at the 7th HAOMO AI DAY, Haomo Zhixing officially launched the MANA OASIS, an intelligent computing center jointly built by it and ByteDance's Volcano Engine. This is also the first autonomous driving company in China. Intelligent computing center established.

The appearance of "Oasis" has opened up a corner of the battlefield in the era of intelligent computing for us, and made Weimo one of the most interesting autonomous driving companies to compete with Tesla.

1. Self-built intelligent computing center, aiming at data closed loop

This should be the largest intelligent computing center in China’s autonomous driving industry.

Floating point operations can reach 67 billion times per second, storage bandwidth is 2T per second, and communication bandwidth is 800G per second. Millimo officially gave such a set of data when introducing MANA OASIS.

Based on Volcano Engine's rich big data accumulation and underlying technology, MANA OASIS realizes computing, storage, and communication capabilities, allowing data to be transformed into knowledge more quickly.

With this new tool, Feimo Zhixing CEO Gu Weihao announced that it is expected that in the first half of 2024, the implementation of NOH in Feimo cities will reach 100 cities, and fully driverless driving will be achieved in 2025.

The emergence of MANA OASIS has its background.

As perception technology and computing platforms gradually mature and converge, the key factors affecting the implementation of high-end autonomous driving are no longer solving common general cases, but solving "intersection" problems, that is, various types of uncommon but constantly emerging The "long tail problem"

As a science that imitates humans, AI autonomous driving is basically consistent with the logic of human cognition of the world. If you want cars to better understand the world, you need to build more accurate models. However, the establishment of algorithm models is not a one-time solution. Self-driving vehicles will always encounter various unfamiliar scenes during driving.

Therefore, how to efficiently process new scene data on a large scale and quickly optimize algorithm models has become the key to iteration of autonomous driving technology. In other words, building a data-driven closed loop of autonomous driving data and allowing data to flow efficiently is the only way to achieve high-level autonomous driving.

In order to meet the ultra-large computing power data center required to realize data closed loop, MANA OASIS appeared.

Its direct task is to provide super computing power for HaoMo Zhixing's MANA data intelligence system and help MANA's five major models achieve low-cost and efficient iteration.

MANA OASIS has four core capabilities, including a robust system architecture to ensure efficient storage and network communication, data management capabilities, computing power optimization capabilities, and training acceleration capabilities.

Gu Weihao introduced that based on the rich big data accumulation and underlying technology of the Volcano Engine, the computing, storage, and communication capabilities achieved by MANA OASIS can allow the GPU to no longer wait for data, and the data can be converted into knowledge faster. The data, computing power, and training efficiency have also been comprehensively improved.

Specifically, in terms of data management capabilities, in order to give full play to the value of the intelligent computing center and allow the GPU to continue to operate at saturation, after two years of research and development, Haimo established a full set of Data Engine for large-scale training, achieving a 10P increase in data screening speed. Times, tens of billions of small files random read and write latency is less than 500us.

In terms of computing power optimization, Feimo cooperated with Huoshan Engine to deploy Lego high-performance operator library, ByteCCL communication optimization capabilities, and large model training framework. Software and hardware are integrated to optimize computing power to the extreme.

In terms of training efficiency, based on Sparse MoE, through continuous optimization of the training platform, Feimo can realize the training of large models with tens of billions of parameters on a single machine, and the method of sharing expertise across machines, completing the training of large models with hundreds of billions of parameters, and reducing the training cost to one hundred Card week level, training efficiency increased by 100 times.

So, how does MANA OASIS improve the data closed-loop efficiency of Haimo?

2. Help MANA upgrade its five major models to make it lower cost and more efficient.

Based on the MANA OASIS intelligent computing center, the five major models of the MANA data intelligence system have been upgraded.

These five major models include: video self-supervision large model, 3D reconstruction large model, multi-modal large model, dynamic environment large model and human driving self-supervised cognitive large model.

Among them, the application purpose of the video self-supervision large model and the 3D reconstruction large model is to reduce costs, and the application of the other three large models is to improve the efficiency of data flow.

Specifically, Haomo Zhixing's video self-supervised large model is mainly used for data annotation, which can realize 100% automation of Haomo's 4D clip annotation and reduce manual annotation costs by 98%.

3D reconstruction of large models can obtain normal cases at low cost and generate various high-cost corner cases. The data generated are not only better and cheaper than the traditional method of manual explicit modeling and then rendering textures, but also increase NeRF. After generating the data, the perceived error rate can also be reduced by more than 30%.

A large multi-modal mutual supervision model can accurately identify special-shaped obstacles. HaoMo Zhixing introduces lidar as a visual supervision signal and directly uses video data to reason about the universal structural expression of the scene. The detection of universal structures can well complement the existing semantic obstacle detection and effectively improve the autonomous driving system in complex cities. Pass rate under working conditions.

The large dynamic environment model can accurately predict the topological relationship of the road, allowing vehicles to always drive in the correct lane.

Based on the BEV (bird's eye view) feature map (feature map), Haomo Zhixing uses the standard map as the guidance information and uses the autoregressive encoding and decoding network to decode the BEV features into a structured topological point sequence to achieve lane topology. Prediction, so that HaoMo Zhixing can realize real-time inference of road topology under the navigation prompts of standard maps like humans in terms of perception capabilities.

Haomo Zhixing believes that solving the intersection problem will actually solve the NOH problem in most cities. Currently in Baoding and Beijing, Haomo has an accuracy of 95% for 85% of intersection topology inferences. Even very complex and irregular intersections can be accurately predicted at any time.

The large self-supervised cognitive model of human driving can master the driving skills of high-level drivers and make driving decisions smarter.

In exploring "the use of a large amount of human driving data, Hao Mo Zhixing newly introduced users' real takeover data, and at the same time used the RLHF (reinforcement learning from human feedback) idea to first train a reward model to select better driving decisions. Through This method has enabled Hao Mo Zhixing to increase the pass rate by more than 30% in recognized difficult scenarios such as U-turns and roundabouts.

In addition, based on the support of Oasis, MANA's latest vehicle-side perception architecture integrates multiple downstream tasks that were scattered in the past to form a more end-to-end architecture, including traffic lights, local road networks, prediction and other tasks, achieving cross- generation upgrade.

The above means that Haimo's perception ability is stronger and the autonomous driving function has more room to explore.

3. What confidence do you have to achieve fully driverless driving by 2025?

At this AI DAY, Haomo Zhixing officially announced that it will achieve fully driverless driving in 2025.

What is the confidence of Hao Mo Zhixing? The answer given by Gu Weihao is the technical route that emphasizes perception and the technical application of large models.

There is a sequence here, which requires a large amount of data first, and then data processing and application.

Before the Intelligent Computing Center, Haimou Zhixing took the lead in introducing new technologies such as Transform and established a data intelligence system MANA in the form of data closed-loop drive. It has become the core driving force for all product iterations of Haimou Zhixing. So far, its learning time has exceeded 420,000 hours. , the driving experience in the virtual world is equivalent to 55,000 years of driving experience of a human driver.