Nvidia Thor: a 2-petaflop automotive platform that wants to replace all the processors in our cars

Already being fully integrated by Chinese electric car brands, Thor is Nvidia’s new car platform. This solution developing up to two petaflops aims to simplify electronic designs. By replacing the hordes of processors and controllers that usually control the various vehicle functions with a single chip.

Farewell Atlan, hello Thor! While Nvidia presented its new generation “Atlan” automotive platform in April 2021, the American revised its copy to replace the old chip with “Thor”. Not only is the deployment of the first vehicles still planned for 2025, but in addition Thor has a definite advantage: instead of the planned 1000 teraflops, the god of thunder will deploy no less than 2000 teraflops.

Two petaflops of power distributed between a CPU, a GPU and Tensor Cores. Dedicated to AI, the Tensor Cores of this new generation of chip are of a new type called “transformers” (inference transform in English). This deep neural network has a key advantage: it learns without having been specifically programmed. Bringing a welcome form of adaptability to the automotive world, where not all scenarios can be written.

Read also: Mobileye introduces EyeQ Ultra, the chip that promises true Level 4 autonomous driving in 2025

Mixing CPU, GPU and AI processor, Thor benefits from all the improvements introduced by the Grace (CPU) and Hopper (GPU) architectures. But also AI-oriented optimizations, such as support for 8-bit floating point calculations (FP8). Between the complexity of 32 and 16-bit models (FP32 and FP16) and the limits of 8-bit integer calculations (Int8), FP8 models are becoming the “happy medium” of AI calculations. And this is where Thor should shine. Bringing all the power that a modern system needs not only to analyze information from many cameras, lidars, etc. But also to control all the technological elements of a car. With a single chip.

A computer to replace them all (and in the AI ​​to drive everything)

As Nvidia’s press release reminds us, there are currently very many (sometimes several dozen) controllers controlling different functions in a vehicle. Creating redundancy certainly, but complicating the design of automobiles – all these little brains have to communicate and synchronize, the supply chain is more complex, etc.

Thor’s 2 petaflops of power allow him to replace all of the vehicle’s various controllers (braking, window control, camera analysis, etc.). If the chip itself will perhaps be more expensive than all the small chips it replaces, its ability to drive several operating systems within should allow the simplification of the electronic design of vehicles. It would indeed reduce the complexity of electrical diagrams, the number of failure points, etc. Not to mention the weight gain (chips, motherboards, cables, protections, etc.) which is an important factor in reducing energy consumption. We are therefore entitled to expect, if not a drop in prices, at least at a constant price. Lowering the cost of developing a platform can limit the cost of more expensive components.

Thor should also be so powerful that it can be entirely dedicated to the piloting part, in particular autonomous. Or, in the case of non-autonomous vehicles, shared between the basic building blocks of driving on one side and the control of the vehicle’s interfaces (driver’s display, infotainmentetc.) on the other.

China and the USA first on the platform

Nvidia’s press release obviously talks about partner car manufacturers… but all of them are Chinese. From Zeekr, which will be the first to launch a vehicle in 2025, to xPeng, the only brands mentioned are Chinese. And the majority of tech companies related to self-driving – Qcraft AI, Aeva, Aeye, Cerence, are American. What about Europeans?

Read also: Mobileye has designed a photonic lidar to launch the autonomous car revolution

According to Danny Shapiro, Vice-President at Nvidia and in charge of the automotive division to whom we were able to ask questions, “dcar companies are of course interested, but I can’t announce anything at the moment “. While it is obvious that the European auto giants are studying all the solutions that come onto the market, this Sino-American overrepresentation sums up well the dynamism of the two “blocks” in the field of electric and “intelligent” cars – just look at the Tesla’s success or read our duel between the Chinese MG4 against the Renault Megane E-Tech Electric to be convinced.

Like all automotive platform announcements, the Nvidia DRIVE Thor platform will not arrive in vehicles tomorrow. The first cars to take advantage of it will leave the factories of the Chinese Zeeker in the course of the year 2025.

Leave a Comment