Meet the Jetson AGX Thor, a $3500 supercomputer for humanoid robots from NVIDIA.

NVIDIA has officially unveiled the Jetson AGX Thor, a compute module for humanoid robotics that promises to set a new pace in the era of physical AI. The developers call it a "supercomputer for robots", and for good reason: the module with 128 GB of memory based on Blackwell architecture produces up to 2,070 FP4 teraflops with power consumption of only 130 watts.

Here's What We Know

The main difference between Jetson Thor and the previous generation(Jetson Orin) is a 7.5-fold increase in power and a 3.5-fold increase in energy efficiency. This leap allows robots to run multiple generative AI models simultaneously and finally respond to the world in real time - be it logistics, agriculture, medicine or humanoid robots.

NVIDIA claims that more than 2 million developers are working on the Jetson ecosystem. Early customers of the new system include Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Medtronic and Meta. And such giants as John Deere and OpenAI are still testing Thor, selecting optimal application scenarios.

NVIDIA CEO Jensen Huang said Jetson Thor is designed for the millions of engineers who are "building robots that can interact and shape the physical world." The system supports popular AI frameworks and integrates with NVIDIA Isaac, Metropolis and Holoscan.

The Jetson AGX Thor Developer Kit is priced from $3500. Production Jetson T5000 modules are already available from global distributors, and production systems can be ordered through NVIDIA partners.

It looks like we're in for a whole generation of robots that can react faster than a human can give a simple command like, "Okay Google, turn on the lights."

Source: NVIDIA