Nvidia Wants to Be the Brains of Your Self-Driving Car

Chip designer Nvidia on Tuesday revealed a new processor called Drive Thor it expects will power the autonomous vehicle revolution. 

Thor processors should arrive in 2024 for cars hitting the roads in 2025, starting with Chinese carmaker Zeekr‘s 001 EV, said Danny Shapiro, vice president of Nvidia’s automotive work. They’re based on Nvidia’s new Hopper graphics processing unit to better handle the artificial intelligence software that’s key to self-driving cars.

“It absolutely will scale up to full autonomy,” Shapiro said, referring to Level 4 or Level 5 self-driving abilities, in which cars can pilot themselves without human occupants paying attention or even present.

Nvidia had planned a chip called Atlan for 2024 but canceled it in favor of Thor, which handles AI software at 2 quadrillion operations per second — twice the speed planned for Atlan and eight times that of its current Orin processor. Thor incorporates one key Hopper feature: the ability to accelerate a powerful AI technique called transformers. Nvidia also expects lower-end Thor variations for the less revolutionary driver-assist technologies like lane keeping and automatic emergency braking.

The automotive processor market is big and getting bigger as carmakers demand more and more processors and other semiconductor chips for driver assistance, infotainment, and the electronic control units that oversee everything from engine combustion to GPS navigation. Each Porsche Taycan has 8,000 semiconductor elements.

Chip designers are cashing in on the new market. Nvidia has $11 billion in automotive chip orders, and a top rival, Qualcomm, has $19 billion in automotive orders in the pipeline.

Also new at Nvidia’s GTC

Among other Nvidia developments at its GTC event:

• Its GeForce RTX 4090 graphics cards, powered by its new Ada Lovelace generation of GPUs for gaming PCs and workstations, will go on sale in October with prices ranging from $899 to $1,599.
• The Jetson Orin line of processors for robots now includes Nano models for smaller robots. They consume between 5 and 15 watts of power for better battery life, cost $199 and up, and start shipping in January. Newly announced companies using Jetson Orin include Canon, John Deere, Microsoft Azure, Teradyne and TK Elevator, Nvidia said.
• The new new Nemo LLM technology is designed to help researchers get more use out of large language models, a hot new area in AI that’s responsible for rapid advances in processing language, imagery and other data. Retraining an LLM consumes massive resources, but the Nemo technology lets researchers perform a much faster incremental AI training that customizes the big AI.

Thor automotive AI chip details

With 77 billion transistors, Thor will be massive, if not the biggest processor on the market. But it’ll let automakers replace a heavier, more expensive and more power hungry collection of smaller chips, Nvidia says. In addition to using Hopper GPUs, it borrows CPU cores from Nvidia’s 2023 Grace processor for conventional computing tasks. It also draws technology from Nvidia’s newest GPU technology for gaming and design, the Ada Lovelace architecture.

The design will make it easier for carmakers to improve their car software with over-the-air updates, Huang said. Tesla has had a big technological lead in that technology for years.

Thor also will be used for robots and medical equipment, Huang said. And it will be able to run three operating systems simultaneously — Linux, QNX, and Android — for different parts of the car computing environment. Partitioning technology ensures the less important work, like infotainment, doesn’t interrupt the crucial safety-related work, Nvidia said.

With autonomous vehicles, promised for years but still only in testing, those chips become even more important.

“The industry has recognized that it’s a much more complex task than initially thought,” Shapiro said of autonomous vehicles. “With safety being paramount, nobody is ready to release these vehicles into the wild until there’s more compute.”