NVIDIA Comprehensive Analysis (VIII), Three Closed Loops, End-to-End Best Practices under Software-Defined Vehicles

This is the 387th original content of "Autobot Reference"

"Empowering Smart Electric Vehicle Organizations and Individuals"

This article is the eighth article of NVIDIA's comprehensive analysis and a summary of the previous seven articles, which mainly introduceSVIDIA's intelligent car solution Drive, and explores how to implement the end-to-end, closed-loop, software-defined car, and fashionable vocabulary of autonomous driving into industry best practices.

For more analysis of Nvidia, please pay attention to the follow-up update of this public account (auto_refer).

A true end-to-end solution

NVIDIA's Drive series is a solution directly for automotive customers and consists of four product pillars, the autonomous driving home development platform Drive Hyperion, the autonomous driving modular software stack Drive SDK, the simulation platform Drive Sim, and the deep learning training platform Drive DGX.

These four product families support each other to form a true end-to-end development process under a unified computing architecture.

Starting from DGX, model training and optimization of autonomous driving perception, planning, and control; Drive Sim verifies the model and algorithm through the simulation data from virtual simulation and the real data collected by the sensor; drive SDK is more like a software shelf, providing basic software, middleware, application software full-stack software; and Hyperion is a suite of data acquisition and verification development. Customers use the SDK and Hyperion to quickly build the technical capabilities of autonomous driving, enabling multi-sensor data acquisition, model training, testing and verification.

Drive Hypersion

Hyperion positions the development platform and reference architecture for autonomous driving, with the goal of enabling customers to quickly build, validate, and deploy L2+ autonomous driving technology, including sensor suites and computing platformS AGX.

The sensor kit includes millimeter-wave radar, lidar, cameras, IMUs, GNSS and other sensors; the AGX Development Kit (Developer Kit) is built on the Orin SoC and provides hardware, software and application services, including the DRIVE AGX Orin, DRIVE AGX Pegasus (L4/L5 Autonomous Driving System) and DRIVE Hyperion 8.1 (Reference Architecture) development kit.

Drive SDK

NVIDIA's DRIVE Software Development Kit (SDK) is a software package for the development of autonomous vehicles, mainly composed of AV and IX application software, middleware DriveWorks, DriveOS toolkit, providing a complete development environment for application software including perception, positioning, mapping, planning control, driver monitoring, and natural language processing.

DriveOS is a basic software stack consisting of RTOS, Hypervisor (QoS), CUDA, TensorRT and other modules, providing application execution and real-time environments such as bootbot, service service, firewall firewall and OTA, RTOS, AUTOSAR and Hypervisor to meet ASIL-D requirements.

DriveWorks is a middleware framework consisting of software module libraries, applications, and related toolchains for open and modular software development through the DRIVE AGX computing platform.

The entire SDK can realize all the functions of active safety, highway driving, urban driving, parking, and cockpit, as shown in the following figure:

Modular design, software and hardware decoupling are the biggest highlights of the Drive SDK, and the key to building NVIDIA's automotive ecosystem.

Customers can choose different modules for development according to their own needs, similar to choosing from the software shelf, when using these software modules, they will naturally be pulled to NVIDIA's hardware, which is NVIDIA's awesomeness.

Drive SIM

DRIVE Sim is a cloud computing platform combined with NVIDIA Metaverse, which mainly provides simulation scenarios such as weather, roads, vehicles, traffic, and virtual worlds for autonomous driving development and verification, and tests and verifies AI algorithms through hardware-in-the-loop (HIL, Hardware in Loop) methods.

Drive DGX

The DRIVE AGX platform is mainly to provide a high-performance end-of-vehicle AI computing platform for autonomous vehicles. The deep learning network model and algorithm after simulation testing can be deployed on the DRIVE AGX vehicle-side platform for corresponding road testing and verification of autonomous driving functions.

In addition, on top of the DRIVE AGX vehicle terminal platform, it is also possible to create or draw a world model and display a 3D surround model of the current vehicle, and autonomous vehicles can also collect sensor data while road test verification.

Therefore, through DGX, large-scale neural networks are quickly verified and trained, and data acquisition, data annotation, data training, simulation simulation, and autonomous driving road test verification are realized, forming a data closed loop.

Autobot Reference Summary

Auto people refer to the Drive product series to achieve vertical, horizontal, depth of the three major closed loops:

Vertically realizes the closed loop from the car end to the cloud, from application software, middleware, basic OS, and then to the underlying hardware calculation; horizontally realizes the closed loop from perception to regulation, from data acquisition, annotation, training, simulation, and verification; and realizes the closed loop of the development process from demand to development to delivery to maintenance, and the whole life cycle of the product.

Therefore, Nvidia provides a rare full-stack closed-loop product for autonomous driving, which is the true sense of the software-defined car.

Nvidia comprehensive analysis of the article summary

Basic information: NVIDIA comprehensive analysis (1), computing platforms bloom everywhere, cars are a dime

GPU Architecture: NVIDIA Comprehensive Analysis (2): Ten years of grinding a sword, GPU and CUDA two wings fly together

Model Quantization: NVIDIA Comprehensive Analysis (3): Deep Learning Model Quantization, TensorRT to understand

Drive platform: NVIDIA comprehensive analysis (4), high-computing power Drive platform to seize smart cars

Xavier SoC: Nvidia Comprehensive Analysis (V), Star Chip Xavier, the most complex and powerful SoC in history

Orin SoC: Nvidia comprehensive analysis (seven), the strongest on the surface, car companies crazy rush, detailed chip Orin

This article is the 387th original article for automotive reference, and if you think the article is good, "recommendation and attention" is the biggest support for me.