Artificial intelligence

NVIDIA Introduces DRIVE Thor, AI Performance Balancing for Full Vehicle Autonomy

Jensen Huang, Founder and CEO of NVIDIA, unveiled the NVIDIA DRIVE Thor, a vastly superior superchip for achieving in-vehicle computing autonomy at a faster rate. The all-new vehicle-grade SoC, or system-on-chip, is built on the latest CPU and GPU advancements to deliver 2,000 teraflops of performance while keeping overall system costs down.

DRIVE Thor is the successor to the company’s NVIDIA DRIVE Orin, featuring the latest computing technology to accelerate the industrial deployment of smart vehicle technology, focusing on automakers’ 2025 models.

The first customer for NVIDIA’s new next-gen platform will be ZEEKR, the Geely-owned premium electric vehicle maker. The company will start production from 2025.

NVIDIA’s DRIVE Thor combines traditionally dispersed vehicle processes including digital clusters, infotainment, parking and assisted driving. The company anticipates greater efficiency in development and faster application iteration.

The DRIVE Thor superchip can be configured by manufacturers in several ways, such as dedicating all 2,000 teraflops of the platform to the autonomous driving pipeline. The new platform will also allow one shard to be used for AI and in-cab infotainment and another to assist drivers.

Like the NVIDIA DRIVE Orin, DRIVE Thor uses the productivity of the NVIDIA DRIVE Software Development Kit, designed to be functionally safe ASIL-D. DRIVE Thor is built on a scalable architecture, allowing developers to seamlessly transfer their previous software development to the next-generation platform.

Additionally, DRIVE Thor produces an extraordinary leap in deep neural network accuracy as well as raw performance.

DRIVE Thor honors the first inclusion of a transformer motor in the family of autonomous vehicle platforms. The NVIDIA-created Transformer Engine is a new component of the company’s Tensor Core GPU. Transformer networks treat video data as a single perceptual image, challenging the computing platform to process additional data over time.

Using FP8, or 8-bit floating point precision, the system-on-chip introduces a new type of data for the automotive industry. Autonomous vehicle developers are seeing a loss of precision when transferring from 32-bit floating-point data formats to 8-bit integer data formats. FP8 precision reduces the difficulty of this transition, ensuring that developers can move data types around without sacrificing precision.

NVIDIA DRIVE Thor uses updated ARM Poseidon AE cores and in turn makes the processor one of the highest performing in the industry.

The system-on-chip is adept at computing multiple domains, dispatching tasks for autonomous driving and in-car infotainment. Multi-compute domain isolation allows concurrent critical processes to run uninterrupted. The vehicle can simultaneously run Linux, QNX and Android operating systems on a single computer. These functions are regulated by dozens of electronic control units scattered throughout a vehicle. Instead of relying on these distributed ECUs, manufacturers can downsize vehicle functions directly using DRIVE Thor’s ability to separate specific tasks.

Everything from vehicle displays to sensors can connect to a single SoC, streamlining automakers’ extremely complicated supply chain. Customers can use one DRIVE Thor SoC or combine two via the latest NVLink-C2C chip interconnect technology to operate as a monolithic platform that leverages a single operating system.

This capability provides automakers with the computing space and flexibility to build software-defined vehicles that can be constantly upgraded through secure, over-the-air updates.

Leave a Reply