Jetson AGX Thor
The NVIDIA Jetson AGX Thor is the new powerhouse in the Jetson family lineup. We’re taking a slightly different approach. Usually when we get a new Jetson, we post a whole bunch of pictures. However, Thor is a different kind of release. It’s sleek and the interface is minimal. That said, it’s also extremely powerful too. Looky here:
Background
Let’s talk in comparison to the last generation of Jetson, the AGX Orin. Usually when there’s a new generation of NVIDIA graphics chips, there’s a new Jetson. However, the AGX Thor skipped over a generation of GPUs. The AGX Orin uses Ampere architecture, the AGX Thor moves to Blackwell.
Physically, the Jetson Thor chip is much larger than the Orin chip. While it still fits on the same size carrier board as the Orin, there’s been a deliberate change to the interfaces offered in comparison.
Specifications
If you’re interested in the specs, take a look at the NVIDIA Jetson AGX Thor Product Page: https://nvda.ws/4mW7wcW
The highlights are:
- 128GB of LPDDR5x memory at 273 GB/s
- 2560 CUDA capable cores
- 96 Tensor Cores
- 1 TB NVMe SSD. The slot is PCIe Gen 5 compatible.
- Operating range of 40W – 130W
Interfaces
The Thor is designed for serious applications, so there are some very high speed interfaces. A new addition is QSFP28 Ethernet, which allows for 4x 25GbpE. That should make it much easier to handle high-speed camera and LiDAR sensor data. The RJ45 interface is 5GbE, which is a nice step up.
Because the Thor is designed for robotic applications, there is now a Microfit power jack. This is much easier to secure than USB-C power. The kit includes a USB-C power supply, which normally powers the Thor when used as a desktop.
But some friends have left
While the Thor gained a monster with the QFSP28, a few options from the previous AGX generations were removed. One is the PCIe slot, and another is the 40 pin GPIO header. This is most likely because of lack of space on the carrier board due to the larger chip size.
Big Boys Need Big Cooling
The AGX Thor is much larger than the AGX Orin, though they have the same size carrier board. This is due to the Thor’s thermal solution. Half of the AGX Thor holds the carrier board and Jetson module. Connected to the Jetson module is a heat pipe which extends to a large fan. The fan takes up the other half of the Thor package.
The area under the fan is empty, and therefore feels less heavy than you expect when first picking it up. The AGX Thor is a sleek, industrial design and sits quite attractively on the desk.
Performance
When NVIDIA provides the test units, they arrive with some nice demos. This time is no exception. However, instead of showing the demos let’s take a look at some tests and comparisons. This gives us a feel for how the Thor performs. One thing to remember is that the Thor software is in beta and will be for a few months, so performance will be better over time.
From previous Jetsons, we usually see an increase of 75-100 % improvement from first release during the first year or two after introduction. This is for a couple of reasons, but in Thor’s case there are other factors to consider. First, Thor is being introduced at a major release of CUDA, version 13. At the same time, the software architecture is changing to SBSA (Server Based System Architecture). That means that the Thor will be able to run the same software as the rest of the NVIDIA ecosystem. For many years, Jetson was the odd duck labeled “Tegra”, which required workarounds. It will take some time to explore its true capabilities and discover what is possible. Let’s take a look at the benchmark and comparison videos:
Conclusion
The Thor is quite the machine. It will take some time to explore its true capabilities and see just how far it can be pushed. NVIDIA is aiming the AGX Thor squarely at the humanoid robotics market, but it’s easy to imagine developers finding all sorts of clever applications far beyond that. With its powerful Blackwell GPU architecture, high-speed interfaces, and a shift toward SBSA compatibility, Thor represents not just another Jetson release, but the start of a new era in edge AI computing.
4 Responses
Bummed about the loss of GPIO. What’s the recommended way to interface the Thor devkit with steppers, servos, etc?
Would love to see some how-to videos on interfacing Thor with the physical world.
I think there are a couple of ways to think about it. The easiest way is to connect to a microcontroller via USB. The microcontroller world is very adept at handling what I’ll call lower level, real time sensors. Typically for something like a servo or stepper, you need some type of interface that provides not only the control signals but also power to the motors.
For stepper motors, people usually use a stepper controller of some sort in any case. For a servo, it depends on the type of servo. If you can run them on a CAN bus, the Thor has two of those built in. Others have some sort of serial bus, so that usually requires a serial bus which can go through the USB ports.
If you’re developing for a product, there are several third party boards with different functions available, depending on what fits your design.
It’s not a satisfying answer, but it’s usually how it would be done in an industrial setting anyway. Thanks for reading!
Now, if only it were affordable for standard humans…
That sounds about right, the price point seems to be oriented towards professional users. Thanks for reading!