NVIDIA Advances On-Device AI Capabilities
NVIDIA's platforms are enabling more powerful AI applications at the industrial edge. The Jetson AGX Xavier platform continues to be a key component for real-time neural network inference in robotics without constant cloud connectivity. Meanwhile, the new DRIVE Thor system raises the performance bar for autonomous vehicle and logistics use cases by delivering 2,000 TOPS of processing power.
- The Jetson AGX Xavier Industrial module is specifically ruggedized for harsh environments, capable of operating in temperatures from -40°C to 85°C and withstanding significant shock and vibration. It includes a "safety cluster engine" with two Arm Cortex-R5 cores in lockstep for error correction and safety functions. - DRIVE Thor represents a significant architectural shift by consolidating functions like automated driving, in-vehicle infotainment (IVI), and parking onto a single system-on-a-chip (SoC). This multi-domain computing capability allows concurrent, uninterrupted operation of critical processes and can run operating systems like Linux, QNX, and Android simultaneously. - A key innovation in DRIVE Thor is the inclusion of a transformer engine, a new component of the NVIDIA GPU Tensor Core. This engine, combined with 8-bit floating point (FP8) precision, accelerates the inference performance of transformer deep neural networks by up to 9 times, which is crucial for the complex AI workloads in autonomous driving. - Compared to its predecessor, DRIVE Orin, which delivers 254 TOPS, a single DRIVE Thor chip provides 2,000 TOPS, an almost eightfold increase in performance. This enables the system to handle the massive data inputs from a vehicle's full sensor suite, including cameras, lidar, and radar, for more advanced autonomous capabilities. - The Jetson AGX Xavier module features a 512-core NVIDIA Volta GPU with 64 Tensor Cores and an 8-core ARM v8.2 64-bit CPU, delivering up to 32 TOPS for AI performance. It is supported by the NVIDIA JetPack SDK, which includes libraries for deep learning, computer vision, and GPU computing to streamline development. - NVIDIA's software platforms, such as Isaac for robotics and Metropolis for vision AI, are designed to work with the Jetson hardware. The Isaac SDK provides tools, a robotics engine, and a virtual simulation environment (Isaac Sim) for developing and training AI-based robots before deployment. - The Metropolis framework simplifies the creation of vision AI applications by providing microservices and APIs for developers. This platform is used across industries like retail and manufacturing for applications such as automated visual inspection and analyzing customer behavior to optimize store layouts. - Automaker ZEEKR, owned by Geely, is the first announced customer for DRIVE Thor, planning to integrate the centralized computer in its intelligent electric vehicles scheduled for production starting in 2025.
