At the spring 2021 Nvidia GPU Technology Conference (GTC) one of the highlights of CEO Jensen Huang’s presentation was the announcement of a next-generation system-on-a-chip (SoC) for automated vehicles called Atlan. Atlan was scheduled to be available for production vehicle applications in 2025. At this week’s fall 2022 GTC, Huang announced that Atlan had been canceled and replaced with a new design dubbed Thor that will offer twice the performance and data throughput, still arriving in 2025.
When it was announced, Atlan promised the highest performance of any automotive SoC to date with up to 1,000 trillion operations per second (TOPS) of integer computing capability. That’s about four times the performance of the Orin SoC that is launching this year in production vehicles including the Nio ET7, Xpeng G9 and the soon to be released Polestar 3 and Volvo XC90 replacement.
At the same GTC where Huang announced Atlan, he also revealed Grace, a new ARM-based CPU for servers. At the April 2022 GTC Nvidia also announced Hopper, it’s next-generation GPU architecture. Given the 2025 timing of Atlan, Nvidia decided that it had time to start over and deliver a new chip that combined technology from the Grace and Hopper chips to create Thor.
The Thor SoC is expected to deliver 2,000 TOPS of integer computing power as well as 2,000 teraflops of floating point performance from 77 billion transistors. For comparison, the Parker SoC that powered version 2 of Tesla AutoPilot (in combination with a Pascal GPU) from 2016 delivered about 1 TOPS and was followed in 2020 by the Xavier chip with 30 TOPs. The Xavier SoC is used in the Xpeng P7 and a number of other vehicles in China.
In addition to the new CPU and GPU cores as well next-generation GPU cores, Thor also integrates NVLINK connections originally developed for data center applications to speed up data transfer between chips on the board. The new SoC is the first automated vehicle compute platform with an integrated inference transformer engine. Essentially, this improves the capability of deep neural networks processing sensor data by running many parallel operations so that the system has context about what is happening at any point in time. With new vehicles now including upwards of 30 sensors including 10 or more cameras, multiple radars, lidar and ultrasonic sensors, this ability is essential making a software perception system work.
This is a chip that is targeted beyond just powering automated driving systems, but also being a full central compute platform for the vehicle. As a central computer, Thor will be expected to manage multiple domains including body control, powertrain, infotainment and driving assist. Nvidia has built in domain isolation to allow the chip to run a variety of operating systems simultaneously including Linux, QNX and Android in protected regions. In addition to the performance improvements, Nvidia is claiming better power efficiency with Thor, although it hasn’t revealed any details yet.
The first production applications with the Thor SoC should arrive sometime in 2025.