Tesla has pioneered the autonomous driving 3.0 era, and Haomo is most likely to be the first Chinese company to do the same, analyst says.

When Xpeng (NYSE: XPEV) announced in early August that it had built what was then China's largest autonomous driving computing center, it probably didn't anticipate that an unknown local startup would build a similar, more powerful facility five months later.

Beijing-based, Great Wall Motor-backed startup Haomo.AI Technology Co Ltd unveiled MANA OASIS, China's largest self-driving computing center to date, at its seventh AI Day event today.

The computing center, launched by Haomo together with Volcano Engine, a cloud service platform owned by ByteDance, has a total computing power of 670 PFLOPS.

For the city assisted driving function, Haomo's City NOH software has reached delivery status and is currently undergoing road tests in several cities.

Models equipped with Haomo City NOH will be available in 2023, and the number of cities supporting the feature is expected to reach 100 by the first half of 2024, the company said at today's event.

Another company that is equally aggressive in building advanced assisted driving capabilities is Xpeng, which said a month ago that it plans to launch its next-generation intelligent assisted driving system, XNGP, in the third quarter of 2023 and the feature will be available in dozens of cities.

The basis for Haomo and Xpeng's ability to be so aggressive in their use of cutting-edge technology is their arms race in infrastructure, and the two companies are currently the only ones with autonomous driving computing centers in China.

Haomo's MANA OASIS has computing power almost entirely for autonomous driving, and its architecture is specifically arranged according to the business needs of autonomous driving.

With MANA OASIS, Haomo has updated its five key models for developing autonomous driving, namely the visual self-monitoring model, the 3D rebuilding model, the multi-modal mutual supervision model, the dynamic environment model, and the driver self-supervision model.

The visual self-monitoring model makes 100 percent automatic annotation of 4D video clips a reality, reducing the cost of manual annotation by 98 percent, according to the company.

The 3D rebuilding model leverages on the NeRF technology to generate highly realistic data by changing the angle of view, illumination and texture materials, allowing Haomo to easily gain data on corner cases, which could be very expensive to acquire otherwise and reducing the rate of wrong perception by 30 percent.

The multi-modal mutual supervision model helps a vehicle to recognize barriers with abnormal sizes, while the dynamic environment model keeps the vehicle always on the right path and the driver self-supervision model learns from experienced and skillful drivers and allows the vehicle to make smarter decisions about how to drive.

In terms of training efficiency, MANA OASIS can complete foundation model training with trillions of parameters, which improves training efficiency by 100 times.

"With abundant data and computing power enabled by MANA OASIS, Haomo's product capability will be even stronger, steering the company into the era of autonomous driving 3.0," said Gu Weihao, CEO of Haomo.

Gu considers the autonomous driving 1.0 era as hardware-driven, the 2.0 era software-driven and the 3.0 era data-driven.

In the 3.0 era, data will be in self-training mode, and self-driving mileage will increase from millions of kilometers in the first two eras to the 100 million kilometer level, said Bai Yu, chief analyst of the auto industry at local Chinese brokerage Pacific Securities, in a December 27 research note.

Tesla has taken the lead in the autonomous driving 3.0 era, while Haomo is most likely to be the first Chinese company to move into the autonomous driving 3.0 era, according to Pacific Securities.

Haomo's autonomous driving data intelligence system MANA has more than 310,000 hours of learning, and the virtual world driving experience is equivalent to 40,000 years of driving experience for human drivers, according to the analyst.

Relying on the advantages of massive data scale, the supercomputing center will further build and strengthen Haomo's competitive advantages and industry barriers, according to Pacific Securities.