Skip to main content

New Details on NVIDIA Jetson AGX Thor: The Future of AI Robotics and Edge Computing

Screen shot showing JETSON Modules Roadmap

NVIDIA has once again raised the bar for AI and robotics with the announcement of the Jetson AGX Thor, a next-generation module designed to power humanoid robots and advanced edge devices. With the Thor module, NVIDIA aims to offer unparalleled performance, leveraging the latest in GPU and CPU architectures.

A Leap Beyond Jetson Orin

JETSON AGX THOR Specification

The Jetson AGX Thor is based on NVIDIA's Blackwell architecture and presents a significant upgrade over its predecessor, the Jetson AGX Orin. Thor offers an astounding 8X performance increase in GPU computation, thanks to the inclusion of a Transformer Engine that supports floating-point precision from FP4 to FP16. This feature is ideal for machine learning models that need rapid and efficient computation.

But the enhancements don’t stop at the GPU. Thor's CPU boasts a 2.6X improvement over Orin. This will allow developers to run more complex models and manage more data streams simultaneously, and less time to compile libraries. Additionally, Thor doubles the memory capacity of the Orin, now supporting up to 128 GB of RAM, essential for data-heavy applications like AI inference and real-time video processing.

The IO capabilities have also received a substantial boost, with 10 times the bandwidth compared to Orin,  this assumes that it will take advanced USB 4.0 support. This increase is complemented by the inclusion new Wi-Fi and high-speed Ethernet interconnections, ensuring that edge devices can communicate seamlessly in high-bandwidth environments.

Thor is set to be released in 2025, and early demos have already generated excitement in the robotics community.

Comparison: Jetson AGX Thor vs. Jetson AGX Orin

Jetson Orin Modules

The Jetson AGX Orin, which has been a leading choice for developers working on next-generation robotics and AI applications, comes in various models, from the high-performance 64GB variant to the more compact Orin Nano and Orin NX modules. The Orin lineup supports between 20 and 275 TOPs (Tera Operations Per Second), with configurations ranging from 4GB to 64GB of RAM.

Orin modules have already been praised for their scalable architecture and efficiency in running AI workloads across devices with different power budgets, from 10W to 75W. However, Thor's leap in performance places it far ahead of Orin in almost every category.

Where Orin is impressive with its Ampere GPU architecture and 12-core ARM A78 CPUs, Thor steps into the future with Blackwell GPUs and more efficient CPUs. The double memory and CPU capacity make Thor ideal for more demanding AI applications that Orin might struggle with, such as large-scale humanoid robotics or advanced autonomous systems.

The Orin remains a fantastic option for a wide range of applications, offering a high level of performance at competitive pricing, with models like the Orin Nano starting at $199 for entry-level developers. Also, Jetson AGX Orin now has low price for high-end edge device.

Jetson Orin Modules Performance

Conclusion

NVIDIA's Jetson AGX Thor represents the future of AI robotics and edge devices. With its transformative GPU and CPU upgrades, increased memory, and faster I/O capabilities, Thor is positioned to lead the next generation of innovation in robotics, AI, and edge computing. As we look forward to its release in 2025, the Jetson AGX Orin remains a powerful and versatile choice for developers needing scalable AI performance today.

As a PhD student in Generative AI and Robotics at the University of Barcelona (UB), I am passionate about developing and researching intelligent systems that enhance human-robot-object interactions (HROIs). With a strong background in AI, edge computing, and GPU technologies, I have the skills and knowledge to tackle complex and challenging problems in this domain.

Comments