NVIDIA DRIVE Thor is a “Centralized Computer” for Autonomous Vehicles
At NVIDIA's most recent GTC, the company announced a new SoC to centralize the disparate computing requirements of modern vehicles.
Between the electrification of vehicles and the push for greater vehicle autonomy, the automotive industry has been one of the major drivers of technological innovation in the past decade. NVIDIA displayed such innovation at its recent GPU Technology Conference (GTC), where the company announced a new SoC designed to centralize many computing requirements within modern vehicles.
NVIDIA DRIVE Thor. Image courtesy of NVIDIA
In this article, we’ll discuss some of the computing challenges for modern vehicles and how NVIDIA's DRIVE Thor SoC is built to overcome these limitations.
Modern Vehicles and the Swell of Disparate Computing Blocks
A modern vehicle is standardly equipped with functions such as advanced driver-assistance systems (ADAS), V2X connectivity, infotainment systems, security and encryption, and much more. From a system design perspective, this is a challenge because many of these tasks, such as ADAS, are extremely computationally intensive.
Modern vehicles handle a wide variety of computing tasks. Image courtesy of MItchell
Vehicle computing systems consist of many disparate computing blocks to support the myriad of functions within a car. To provide the best performance and power efficiency, designers utilize hardware accelerators and dedicated computing blocks for each task. This scheme is not scalable or cost-effective. And, as cars become more functional in the future, the swell of overburdened computing blocks is only expected to grow.
DRIVE Thor Centralizes Computing
To help address the data demands of the modern vehicle, NVIDIA recently announced a new SoC during the company's GTC keynote address.
The new SoC called the DRIVE Thor is described as a “centralized computer for safe and secure autonomous vehicles.” To this end, NVIDIA has designed Thor to be a single computing engine that unifies many of the currently disparate computing tasks within a vehicle. Specifically, NVIDIA claims that the DRIVE Thor is optimized to simultaneously support tasks such as automated and assisted driving, parking, driver and occupant monitoring, digital instrument cluster, in-vehicle infotainment, and rear-seat entertainment.
To do this, the SoC comes equipped with AI capabilities that originate from the Hopper GPU architecture along with the Grace CPU and Ada Lovelace CPU. Uniquely, NVIDIA also says Thor is the industry’s first vehicle SoC to incorporate an inference transformer engine, a new component of NVIDIA's Tensor Engine. With the inference transformer engine, NVIDIA claims to accelerate transformer deep neural network performance by 9x.
NVIDIA's Hopper architecture employs the company's transformer engine. Image courtesy of NVIDIA
Further, the SoC is capable of multidomain computing, meaning it can partition compute tasks to concurrently run time-critical processes without interruption. Altogether, Thor features a max performance of 2000 TFLOPS of FP8—a marked improvement over its predecessor, the NVIDIA DRIVE Atlan, which maxes at 354 TOPS.
The Design and Market Hope of DRIVE Thor
With DRIVE Thor, NVIDIA seeks to solve one of the greatest system challenges in modern vehicles. By bringing the many disparate computing engines within a vehicle together, DRIVE Thor may reduce BOM, cost, weight, size, and system complexity for the design engineer. Altogether, the hope is that these improvements will lead to more affordable and performant vehicles moving forward.