Nvidia goes after Tesla FSD with Thor Blackwell AI car chip and 4D autonomous driving world simulator

Nvidia used its CES 2025 keynote to announce a new Thor automotive chip based on the Blackwell platform that has 20x the processing power of its current generation of car chips, and can be used in virtual autonomous vehicle scenario simulations.

Instead of logging millions of miles and citing intervention data like Tesla does with its FSD option, Nvidia bets on a new synthetic physical world simulator it calls Cosmos that will be added to the trio of chips used for autonomous vehicles AI.

The Cosmos platform consists of "state-of-the-art generative world foundation models, tokenizers, guardrails and an accelerated video processing pipeline built to advance the development of physical AI systems such as AVs and robots," tips Nvidia.

Nvidia Thor Blackwell car chip specs

The chips responsible for all that AI learning also got an upgrade as Nvidia announced Thor, the successor to its Orin line of car processors. The Thor automotive chip is based on the newfangled Blackwell architecture, just like the new GeForce RTX 5070 gaming card that Nvidia says delivers the performance of an RTX 4090 for just $549.

Thor offers 1,000 TFLOPS of high-powered AI computing and about 20x the AI performance of its predecessor, but what's even more important is that it is more power-efficient and can reduce the overall SoC costs in a combined autonomous driving and infotainment car computer.

Car makers and autonomous vehicle solution developers can decide what proportion of the 1,000 TFLOPS of computing power they will dedicate to self-driving and what to cabin infotainment duties, conveying a highly flexible architecture.

"When deployed in level 3 or higher autonomous driving use cases, a single Nvidia Drive AGX Thor will deliver the same functionality as multiple leading edge devices in vehicles today, leading to significant performance benefits for automakers," tips ARM.

Nvidia also boasted a trifecta of safety and security certifications for its new Thor-based Drive AGX platform, ranging from the ANSI National Accreditation Board to the usual suspects from TÜV Rheinland.

Nvidia has already signed a number of automakers and autonomous driving system providers, like Mercedes or Nuro, as clients for its self-driving chips and Cosmos simulation software whose essence it explains as follows:

OmniMap fuses map and GEO satellite imagery to construct drivable 3D environments. Neural Reconstruction Engine creates high-fidelity 4D simulation environments from AV sensor logs. Using text or image prompts, Edify 3DS can search for existing scenes, or generate 3D objects, transforming scenes in NVIDIA Omniverse that are 4D grounded. Lastly, Cosmos can generate near infinite variations of driving scenarios using text prompts.

It, however, remains to be seen if the synthetic Cosmos simulator will be the magic potion that allows legacy automakers and EV startups alike to catch up to Tesla's FSD solution that relies on millions of real-world miles of driving to train its AI algorithms.