Key Points
- 1.The Jetson Thor devkit features 128 GB of video memory and costs around $3,000.
- 2.It has a memory bandwidth of 273 GB/s, which is fast for CPU/RAM but slow for GPU standards.
- 3.Ideal for robotics, it draws only 130 watts of power and is not meant for training models, only inference.
Summary
Device Overview
The Jetson Thor devkit is highlighted for its impressive 128 GB of video memory and affordable price of approximately $3,000. This devkit is compact, with most of its hardware encapsulated in a small square board designed for easy embedding into robotic applications.
Memory Bandwidth Comparison
With a memory bandwidth of 273 GB/s, the Jetson Thor is considered fast for CPU/RAM but much slower than high-end GPUs like the 5090. This difference indicates that while it's capable for certain applications, it does not compete with dedicated GPU hardware for tasks like model training.
Power Efficiency
One of the standout features of this devkit is its low power consumption of just 130 watts, making it suitable for portable robotic applications. This power efficiency enables it to perform complex computations without significant energy use.
Target Audience and Use Cases
The Jetson Thor is designed particularly for developers and engineers working in robotics, providing the ability to run local language models and other applications without the need for expensive, high-powered computing setups. It's positioned as a niche device focused on inference rather than training models.
Scalability and Limitations
While multiple units can be networked together, the memory bandwidth limitations present a bottleneck for scalability. Users are encouraged to explore its capabilities with local models, emphasizing the unique value of high memory capacity in robotics.
Worth watching for
This video is for robotics developers and engineers interested in the Jetson Thor devkit for local model inference applications.