The "spring" of autonomous driving commercial services is coming.
Yesterday, NVIDIA announced that a group of Chinese autonomous driving startups, including Yuanrong Qixing, Wenyuan Zhixing, and Yunji Zhixing, will join the DRIVE Hyperion ecosystem, equipped with an on-demand customized DRIVE Hyperion8 programmable platform architecture, based on end-to-end solutions to accelerate autonomous driving development.
At the beginning of this year, another Chinese autonomous driving startup, Xiaoma Zhixing, also announced that it would achieve a milestone breakthrough in core computing units from industrial to vehicle specification level based on NVIDIA's vehicle specification-level DRIVE Orin.
Prior to this, NVIDIA has successively won the next-generation computing platform front-loading orders of Mercedes-Benz, Volvo, Jaguar Land Rover, Weilai, Ideal, Xiaopeng, BYD and many other car companies.
How important is the three words "vehicle regulation level" for L4 level automatic driving?
According to the official disclosure of Xiaoma Zhixing, the mass production of the next generation of autonomous driving software and hardware systems means accelerating the large-scale deployment of L4 autonomous driving technology. Prior to this, most L4 autonomous driving companies developed systems based on industrial-grade or front-scale mass-produced computing platforms, mainly for the purpose of prototyping and testing the system.
Yuanrong Qixing also believes that the integration of DRIVE Hyperion into its L4 level autonomous driving system means that a complete automotive-grade platform will be the key to the company's mass production L4 autonomous driving plan to the market.
Of course, similar front-loading computing platform solutions, Nvidia is not the only one. For example, based on Horizon's latest generation of high-computing power vehicle specification-level AI chip - Journey 5, the vehicle computing platform Matrix5 is fully adapted to Horizon's leading algorithm software module, the hardware configuration can be customized, open, flexible and easy to use and independent innovation.
One
DRIVE Hyperion, what is it?
This is an end-to-end modular development platform and reference architecture launched by NVIDIA at the end of last year, which can be used for L4 autonomous driving. Includes NVIDIA DRIVE AGX Orin, DRIVE AGX Pegasus, and DRIVE Hyperion Development Kits, and is based on the upcoming NVIDIA DRIVE Orin computing platform.
This scalable platform includes the underlying NVIDIA DRIVESDK-DRIVE OS and driveworks, as well as sensors that provide the highest level of safety and efficiency in autonomous driving development.
Among them, NVIDIA DRIVE Hyperion7.1 is a reference architecture for L2+ level autonomous driver assistance solutions, including a complete sensor suite and AI computing platform, as well as a complete software stack for autonomous driving, driver monitoring and visualization.
NVIDIA DRIVE Hyperion 8.1 is suitable for L4 level autonomous driving platform development, the complete toolkit can be integrated into the test vehicle, developers can quickly develop, evaluate and validate technology. At the same time, based on NVIDIA DRIVE OTA update tools and services, software updates are implemented.
This set of AI calculations based on the high-power car specification grade Orin, a complete sensor set and a complete software stack. At the same time, this solution provides sensor calibration, time synchronization, data compression, virtual simulation (NVIDIA DRIVE Sim) and so on.
Among them, Orin uses six different types of processors, including CPU, GPU, deep learning accelerator (DLA), programmable vision accelerator (PVA), image signal processor (ISP), and stereo/optical flow accelerator, and adopts redundant and diversified strategies to ensure the safe operation of autonomous driving systems.
In addition, Nvidia provides reference designs through cooperation with Sony (cameras), Continental (4D millimeter wave radar, long-range radar), HELLA (short-range millimeter wave radar), Valeo (ultrasonic sensor), Luminar (lidar), U-Blox (GPS), Hesai Technology (truth value system) and many other manufacturers.
Compared to the DRIVE AGX Orin development kit (Where Tier 1 still plays a key role), DRIVE Hyperion actually emphasizes NVIDIA's desire to give downstream application partners a more complete autonomous driving development capability.
Prior to this, most L4 autonomous driving companies were through scattered external development tools, customized non-vehicle-level sensors, and some self-developed capabilities, such as simulation testing, truth-value systems, and so on.
In addition, Nvidia has officially released Drive Map, a multimodal map engine, through DeepMap, a high-precision mapping company acquired last year, and plans to collect 500,000 kilometers of data in North America, Europe and Asia by 2024, providing embedded centimeter-level precision navigation for L3/L4 autonomous driving.
At this point, the Orin/Atlan computing platform, the Drive Map map engine, and the Hyperion development platform mean that NVIDIA is no longer just a traditional automotive chip provider, but a full-stack solution provider across the hardware and software ecosystem.
In addition, yesterday Nvidia also announced a new generation of DRIVE Hyperion architecture based on the next-generation Atlan computing platform, which will enter mass production in 2026, supporting 14 cameras, 9 millimeter-wave radar, 3 lidar and 20 ultrasonic sensors, as well as 3 cameras and a millimeter-wave radar in the cabin.
At the same time, Drive Hyperion enables cross-generational compatibility: it can be seamlessly migrated from Orin to Atlan or later in the future. For end users, this means that as long as they choose the NVIDIA platform, they do not have to worry about the additional software development costs that may be brought about by subsequent hardware updates.
This is also the general trend of the industry.
For example, the Horizon Journey series of chips continues to iterate, and the third-generation automotive-grade AI chip Journey 5 has both large computing power and high performance, and is currently the only domestic autonomous driving chip in the industry that has been verified by front-loading mass production, and the computing performance is better than Orin under the key tasks of automatic driving. In the large computing power autonomous driving chip market, Horizon 5 is also a strong competitor of NVIDIA Orin in the Chinese market.
Two
Performance, cost, flexibility and openness, portability and cost sharing at scale for consumer-grade and mobility services make chipmakers such as NVIDIA build their own "moats" on the industry's optional computing platforms.
According to the information that has been made public, some self-driving startups have set the mass production time of the NVIDIA DRIVE Hyperion platform in the first quarter of 2023. Considering that Nvidia Orin will enter the front-loading passenger car market for private users starting this year, this means that the cost of computing platforms will have the opportunity to achieve a rapid decline.
NVIDIA's order data shows that as of now, the company has received a total of $11 billion in order contracts for a series of software and hardware solutions such as Orin and DRIVE Hyperion, which will be delivered in the next six years. This means that for L4 self-driving companies, there is a visible cost reduction space.
"This progress marks another important step in the commercialization of our L4-level solution." In the view of Zhou Guang, CEO of Yuanrong Qixing, by quickly integrating self-developed inference engines and algorithms into DRIVE Hyperion, vehicle design and verification costs will be greatly reduced.
According to the plan, the DeepRoute-Driver 2.0 platform launched by Yuanrong Will be equipped with two DRIVE Orins (hash rate 508TOPS) to deal with perception, decision-making and deep neural network algorithms at the same time.
On the sensor side, the system, which costs less than $10,000, will feature three ML-30s short-range (large-angle MEMS) lidar from One Path Technology, Sagitar Juchuang's second-generation intelligent solid-state lidar RS-LiDAR-M1 and eight high-definition cameras.
In terms of product planning, from 2022 to 2023, Yuanrong Qixing will expand in-depth technical cooperation with the main engine factory, develop a vehicle-grade front-loading mass production plan, and it is expected that by 2024, cars equipped with L4 automatic driving systems will begin mass production and enter the market on a large scale.
All of this is due to the synergy between the consumer level and the mobility/logistics autonomous driving service.
For example, in Mobileye's view, the relationship between Robotaxi and consumer-level autonomous driving is not conflicting, but synergistic, and it is important to establish this relationship.
"The reason we are now able to afford consumer-grade autonomous driving capabilities is because in our Robotaxi solution, there are very strict design constraints on the cost of the system." Erez Dagan, executive vice president of product and strategy at Mobileye, said.
In fact, whether it is Xiaoma Zhixing or Yuanrong Qixing, these autonomous driving startups are also stepping up in-depth cooperation with car companies. For example, Xiaoma Zhixing's mass production plan for the L4 vehicle specification level is being tested on Toyota's customized version of the Sienna, which is the first mass-produced autonomous vehicle to be delivered from the existing production line.
In turn, the rapid scale advantage of the consumer market has driven the cost of vehicle-grade software and hardware to decline. These costs are still one of the bottlenecks in the current L4 autonomous driving, especially the large-scale deployment of Robotaxi.
For example, last year, Horizon announced that it reached the intention of mass production of the First Series of Journey 5 chips with many automobile manufacturers such as SAIC Motor, Great Wall Motor, Jiangqi Group, Changan Automobile, BYD, Nezha Automobile, and Lantu Automobile, and reached a pre-research cooperation based on Journey 5 with Ideal Automobile, accelerating the popularization of high-level automatic driving functions, and obtaining mass production points from many car companies. At present, Horizon is the only domestic and global only three enterprises to achieve the front-loading mass production of vehicle-grade artificial intelligence chips.
"Consumer-grade autonomous driving is the industry's endgame." In the view of Amnon Shashua, CEO of Mobileye, by developing entire autonomous driving solutions, from hardware and software, from maps to end-to-end models, maximum optimization of performance and cost can be achieved.
In fact, the fork in the road to commercial services such as consumer and travel/logistics is ultimately the same destination. For the current automotive chip manufacturers, the upper limit of capabilities will determine the position of future market share.