As the only hardware released on GTC China 2019, NVIDIA drive AgX Orin light is very bright in terms of hardware parameters, including 17 billion transistors, 8-core 64 bit CPU, and 200tops deep learning computing power. On the stage, Huang Renxun repeatedly emphasized the performance of NVIDIA drive AgX Orin, which is seven times that of the previous generation of NVIDIA drive AgX Xavier, which is a completely innovative processor from the underlying architecture.
So the question is, what kind of changes will NVIDIA drive AgX Orin bring to automatic driving, and what kind of planning does NVIDIA have for the layout of automatic driving in the future? At the GTC China 2019 site, we interviewed Mr. Danny Shapiro, senior director of NVIDIA's automobile business department, and asked him to uncover the mystery about Orin and NVIDIA's automatic driving technology.
Mr. Danny Shapiro, senior director, NVIDIA Automotive
Put software first
In NVIDIA, the number of software developers has actually exceeded the number of hardware engineers. The NVIDIA drive AgX Orin chip launched is more about the NVIDIA drive platform. By providing pre training model, the entire NVIDIA drive ecosystem is opened to all partners.
Huang Renxun holds NVIDIA drive AgX Orin chip, which is the only new hardware product of GTC China 2019
Take didi for example, the cooperation between NVIDIA and didi automatic driving has been in a state of continuous promotion, and the underlying technology based on NVIDIA drive continues to expand. With the help of NVIDIA AI, didinen will soon consolidate and complete L4 level automatic driving and intelligent taxi service.
This will be a huge market. At this stage, Didi provides travel services for 10 billion people every year. Relying on chips and edge computing alone cannot solve the overall problem. Therefore, the NVIDIA drive platform and NVIDIA AI data center can work together to solve the problem. It is an optimal solution to provide didi with a complete solution from the vehicle to the cloud.
As a part of Didi's AI processing, NVIDIA drive integrates data from various sensors (cameras, lidars, radars, etc.) with the help of multiple depth neural networks, so as to achieve a 360 degree all-round understanding of the car's surrounding environment and plan a safe driving path. In order to train these deep neural networks, Didi will use NVIDIA GPU data center server. In terms of cloud computing, Didi will also build Ai infrastructure and launch computing, rendering and game based vgpu cloud servers.
The NVIDIA Drive platform can be adapted to be automatically driven from L2, L2, which is now invested in a soft and hard-to-play scheme, and is also sufficient to prepare for the upcoming L4, L5 auto-driving. It is also because in the same system, while an essential change in structure occurs, Orin can be compatible with Xavier, thanks to the fact that the hardware itself can be defined entirely by software.
Interestingly, NVIDIA's involvement in the field of automobile didn't start from the high-end automatic driving, but from the full LCD dashboard, vehicle entertainment information system, for example, Audi also used tegar series chips to control laser headlights. With the continuous upgrading of demand, NVIDIA began to intervene in automatic driving, and has developed to the present NVIDIA drive platform and NVIDIA AI data center. For NVIDIA, only by relying on data center and learning various possibilities of automatic driving simulation through neural network, can it have a chance to lead high-level automatic driving to mass production faster.
In fact, Tesla's autonomous driving system is similar to NVIDIA's concept. The difference is that NVIDIA's automated driving solutions can accommodate more models, rather than being limited to single brands.
Karma relies on the AI assisted driving system developed by NVIDIA drive AgX platform
Because of this, NVIDIA drive AgX Orin provides a complete set of hardware and software solutions, especially in the software part, NVIDIA is the only one that has submitted test applications for all categories above the mlperf benchmark, and ranks first in all categories. Orin has a complete software stack and supports all mainstream AI frameworks.
Win win with all parties
High precision map is undoubtedly an important indicator of automatic driving. At the GTC conference, NVIDIA drive localization was also supported by high-precision map from Gaud map and wide stool technology. NVIDIA DRIVE Localization itself is an open and extensible platform, and the main purpose is to enable autopilot to achieve centimeter level precise positioning on the global road.
By matching the semantic landmarks in the driving environment with the features in the high-precision maps of companies such as Gaud map and kuanstool technology, NVIDIA drive localization realizes centimeter level positioning, so as to determine the exact location of the vehicle in real time.
From the operational requirements of the auto-driving vehicle, a car is likely to require more than 10 cameras, with more than one radar, and a lidar, which requires redundancy and diversity. The demand for data processing also becomes higher. NVIDIA DRIVE AGX Orin that makes a complete change from the underlying structure can handle these issues with a 7-fold higher performance than Xavier.
In fact, the design of NVIDIA drive AgX Orin itself originated from the data center. It took the characteristics of security and stability into consideration at the beginning of the design, and then relied on the software to continuously define and add new functions. In terms of security and scalability, it has a better guarantee.
NVIDIA expects to see the emergence of self driving taxis operating on designated routes in the second half of 2020. NVIDIA has prepared complete solutions for high-precision maps, autonomous driving platforms and data centers. There is no doubt that this is a technology enterprise driven by GPU, but it will not be limited to GPU. It is believed that in the near future, we will also see more self driving vehicles with NVIDIA chips on the road to test on the actual road.