Nvidia announces Drive Thor SoC to centralize car compute

Nvidia announced Drive Thor, an ambitious centralized car computer cluster to unify functions for automated driving and in-car infotainment to first appear in Zeekr intelligent EVs starting production in 2025.

Thor is a System on Chip (SoC) that replaces Drive Atlan on the same timeline and will be the next generation for Drive Orin, now in production, Nvidia said Tuesday at its GTC event.

“Drive Thor is the superhero of centralized compute,” Nvidia CEO Jensen Huang declared.  Both its superhero name and projected functionality show Nvidia’s ambitions for the evolving smart car market where it already claims superiority. In his keynote, Huang said Thor might find use in industrial and robotics applications beyond vehicles.

“No one else provides this complete compute solution” for modern vehicles, said Danny Shapiro, vice president of automotive for Nvidia, told reporters. Atlan had been projected to reach 1000 TOPS of performance, but Thor doubles that amount to 2000 TOPS, he noted.

In concept, Thor will allow carmakers to plug all their dozens of sensors for lidar, radar, cameras and more along with various systems into the Thor SoC where the data will be sorted through its inference transformer engine, a new part of Tensor Cores within Nvidia GPUs. Inference performance is advertised as being increased by nine times over existing tech.   Thor will include Nvidia Grace CPU and Hopper GPU cores, versions of which Nvidia earlier described as superchips for AI use in data centers.

RELATED: Nvidia ‘superchips’ set to turn data centers into AI factories

The shift to a centralized electronic architecture will be a big one for carmakers, but will reduce the need for cabling and individual electronic control units near the many sensors throughout a vehicle. As a result, weight and cabling will be reduced as well as power needs, at least in concept.

“Nvidia is taking all of its technologies and combining them to try to create a single chip that really can be the central computing hub of the vehicle,” said Sam Abuelsamid, principal analyst for e-mobility at Guidehouse Insights in Detroit.

Ironically, the chip shortage that hit carmakers so hard has prompted them to introduce new compute architectures “that get away from a lot of the legacy chips and the distribute computing model that has built up piecemeal over the past 40 years,” Abuelsamid said via email. Carmakers want to jump to a central compute architecture to do some signal processing and power distribution.

Getting to the new architecture requires a high performance SoC. “Thor seems to fit the bill,” Abuelsamid said. Massive data throughout will be provided through NVlink technology, which came from Nvidia’s Mellanox acquisition.

While Thor is expected to be able to suck up data from ADAS and AD sensors and other systems to run on a central platform, a big question remains about power consumption and  cost, Abuelsamid said. “Automakers want to go this direction, but can they afford to in dollars or watts for EVs?” he asked. 

Nvidia’s Shapiro had a response to questions about affordability, even if he had no specifics to offer on power savings or the cost of the Thor SoC. “Reducing several ECUs [Electronic Control Units] into one will result in tremendous savings in cost and reduced cabling, weight and energy consumption,” Shapiro told reporters.  Thor will be able to process both raw sensors data or preprocessed sensors data, he said.

Also, vehicle customers will be able to get a single software update for multiple improvements, he said. Thor will enable carmakers to scale a vehicle to full autonomy running both AV and IVI apps.

Cepton innovation

Nvidia also announced at GTC its Drive Sim simulation platform is getting a new suite of tools to help designers build future cars. Cepton, a San Jose innovator, announced it will add its lidar models into Drive sim based on its lidar sensors.

 Cepton becomes the first lidar partner to provide both near and long range simulation, the company said in a release. Two-fold detection helps eliminate blind spots and obstacle detection at highway speeds.