Nvidia’s Thor Hammers Atlan

Nvidia has revised its automotive roadmap, replacing the product code-named Atlan with the new Thor processor, which targets an impressive 2,000 trillion operations per second (TOPS).

Linley Gwennap

Nvidia has revised its automotive roadmap, replacing the product code-named Atlan. The new Thor processor targets an impressive 2,000 trillion operations per second (TOPS), twice the performance of Atlan, while still aiming to reach production in early 2025. The company positions it as a single-chip solution for autonomous vehicles: a central processor to safely and securely control all driving functions. But it released few technical details of its next-generation product, which remains in development.

In its quest to power self-driving (Level 4) vehicles, the company has delivered a series of AI-enabled processors along with boards that combine multiple chips. The original Parker chip in 2016 generated only 1.5 TOPS, but Xavier boosted that figure to 30 TOPS two years later. The new Orin chip can generate 127 TOPS for dense networks and 254 TOPS in sparsity mode.

Even this performance, however, is inadequate for Level 4 autonomous driving. To help customers develop their software on prototype vehicles, Nvidia combined a Xavier processor with a Volta V100 GPU to create the Drive Pegasus platform, generating 320 TOPS. The new Drive Pegasus OA version sports two Orin processors and two Ampere A100 GPUs, delivering 700 TOPS without sparsity. Using the new 8-bit floating-point (FP8) data format for inference, the Drive Thor board ups the ante to 1,000 TOPS without sparsity or 2,000 TOPS with sparsity.