This week Jensen Huang, founder and CEO of software company Nvidia, unveiled Drive Thor, the company’s new ‘superchip’ to support autonomous driving.
The automotive-grade system-on-a-chip (SoC) is built on what Nvidia describes as ‘the latest CPU and GPU advances’ and can reportedly deliver 2,000 teraflops of performance, while also reducing overall system costs.
Drive Thor succeeds Nvidia Drive Orin in the company’s product lineup, incorporating newer compute technology to support the acceleration of industry deployment of intelligent-vehicle technology.
The chip was produced to unify traditionally distributed functions in vehicles — including digital cluster, infotainment, parking and assisted driving — for greater efficiency in development and faster software iteration.
Manufacturers can configure the Drive Thor chip in multiple ways, including dedicating all of the platform’s 2,000 teraflops to the autonomous driving pipeline, or using a portion for in-cabin AI and infotainment and another portion for driver assistance.
Like the current iteration of Nvidia Drive Orin, Drive Thor is designed to utilise the productivity of the Nvidia Drive software development kit, to be ASIL D functionally safe and is built on a scalable architecture, meaning developers should be able to seamlessly port their past software development to the new platform.
Drive Thor was also designed to deliver significant advances in deep neural network accuracy and, Nvidia says, marks the first inclusion of a transformer engine in the AV platform product offering.
The transformer engine is a new component of the Nvidia GPU Tensor Core. Transformer networks process video data as a single perception frame, which supports the compute platform to process a growing amount of data over time.
Equipped with 8-bit floating point (FP8) precision, the SoC introduces a new data type for the automotive industry. Traditionally, AV developers see a loss in accuracy when moving from 32-bit floating point to 8-bit integer data formats. However, FP8 precision is designed to support and simplify this transition, making it possible for developers to transfer data types without compromising on accuracy.
The SoC is reportedly capable of multi-domain computing, meaning it can partition tasks for autonomous driving and in-vehicle infotainment. This multi-compute domain isolation lets simultaneous time-critical processes run without being interrupted. This means the vehicle can concurrently run Linux, QNX and Android on a single computer.
Generally, these types of functions are managed by tens of electronic control units (ECUs), which are distributed throughout a vehicle. Rather than relying on these ECUs, manufacturers can now consolidate vehicle functions using Drive Thor’s function of isolating specific tasks.
According to Nvidia, all vehicle displays, sensors and other components are designed to connect to the single SoC, making what has traditionally been a complex supply chain for automakers into a more efficient system. This thus offers auto-manufacturers the compute headroom and flexibility to build software-defined vehicles that can be continuously upgraded via over-the-air updates.
