Nvidia GTC China Conference: Open Source Autonomous Driving AI Algorithms, Establishing Partnership with Didi

2019-12-19

At the annual Nvidia GTC China technology conference, founder Huang Renxun showcased a series of products and services in his keynote speech. In addition to the release of the latest version of Drive AGX Orin, a software-defined solution for autonomous vehicles and robots, Nvidia also announced the open-source AI model suite for autonomous decision-making and visual perception. More importantly, they announced a partnership with Didi to pave the way for new developments in autonomous driving projects. New Autonomous Driving Platform: Drive AGX Orin Nvidia's Drive AGX Orin, set to be released in 2022, will be used in conjunction with the existing AGX Drive platforms, Xavier and Pegasus. Huang Renxun stated that unlike Xavier, designed for highly automated and fully autonomous driving, and Pegasus, which powers advanced driver-assistance systems, Orin is designed from scratch to run multiple applications and models simultaneously while meeting safety standards such as ISO 26262 ASIL-D. The silicon core of Orin is a brand-new system-on-a-chip, containing a total of 17 billion transistors, the result of several billion dollars in research and development over four years. It integrates Nvidia's next-generation graphics chip architecture and the Hercules kernel, based on Arm's DynamIQ technology, improving power and area efficiency by 10% compared to the previous generation. It is complemented by AI and machine learning accelerator cores, capable of delivering up to 2 million trillion operations per second (TOPS), while Pegasus and Xavier offer 320 TOPS and 30 TOPS, respectively. Additionally, Orin can handle over 200 Gbps of data while consuming only 60-70 watts of power (200 TOPS). Like Pegasus, Orin can scale from Level 2 (able to control either steering or speed but not both simultaneously) to Level 5 (fully autonomous driving without supervision). Similar to Xavier, Orin can be programmed using CUDA APIs and libraries, Nvidia's parallel computing platform and application programming interface model, as well as the TensorRT platform, to achieve high-performance machine learning inference. Huang Renxun mentioned that when shipping to manufacturers, Orin, based on a single architecture, will have multiple configurations. Building safe autonomous vehicles is perhaps the biggest computational challenge in society. Open-Source Autonomous Driving AI Algorithms Alongside AGX Drive Orin, Nvidia announced that it will provide open-source access to several models within the Drive stack, a full-stack autonomous driving and driver-assistance solution. Specifically, it plans to offer tailored AI systems that come with Drive AGX, targeting tasks such as traffic signal and sign recognition, vehicle and pedestrian detection, path perception, gaze detection, and gesture recognition. Many of these models have been jointly developed for years and have been widely used by car manufacturers, truck manufacturers, autonomous taxi companies, software companies, and universities. Autonomous vehicles are software-defined cars that require operation with various datasets on a global scale. By providing developers with access to our models and advanced learning tools for optimization across multiple datasets, we can achieve shared learning between companies and countries while maintaining data ownership and privacy. Ultimately, we are accelerating the realization of global autonomous driving. Nvidia's newly released toolkit can be used to customize and enhance each model, utilizing various machine learning development techniques for training. Huang emphasized some of the technologies, such as active learning, which improves accuracy and reduces data collection costs by using AI for automated data selection; federated learning, which allows the use of datasets nationwide and with other parties while preserving data privacy; and transfer learning, which leverages pre-training and fine-tuning to develop models for specific applications and functionalities. Partnership with Didi Lastly, Nvidia revealed that Didi will use Nvidia graphics cards in its data centers to train machine learning algorithms that support its autonomous systems. Didi's vehicles utilize Nvidia's Drive platform to fuse data from multiple sensors, including cameras, lidars, and radars, to ensure a comprehensive understanding of the surrounding environment. Didi also announced the establishment of an AI infrastructure and the launch of virtual GPU cloud servers for computation, rendering, and simulation. Additionally, Didi Cloud, the cloud technology department of Didi, will adopt a new virtual graphics card licensing model to provide Nvidia-driven cloud computing services tailored to traffic, artificial intelligence, graphics rendering, video games, and educational workloads. Rishi Dhall, Vice President of Nvidia's Autonomous Vehicles division, stated, "Developing safe autonomous vehicles requires end-to-end AI implementation in both the cloud and the car. Nvidia AI will enable Didi to develop safer and more efficient transportation systems while providing extensive cloud services."

RECENT AI NEWS

RECENT AI TOOLS