Nvidia’s Jetson AGX Orin Packs an AI Punch in a Small Package
by David Cardinal
If there is any company that’s on a roll when it comes to providing more compute power in smaller packages, it’s Nvidia. Their Jetson product line that provides AI and other forms of accelerated computing is a great example. I’ve been able to spend some time with its latest embeddable “robot brain” offering — the Nvidia Jetson AGX Orin (starts at $399 when available for production applications later this year). It has enough GPU power for some of the most demanding robot applications, while still fitting in the same form factor as Xavier, its predecessor. It consumes from 15 to 60 watts, depending on the power profile used.
What we’re reviewing here is a developer kit ($1,999), that comes complete with an enclosure and some accessories. It is available now (in theory, but back-orders have been piling up) so that developers can get a head start, but volume quantities of the version suited for commercial deployment aren’t expected until near the end of the year.
Nvidia Jetson Orin by the Numbers
The Orin System-on-Chip (SoC) is based on the Nvidia Ampere GPU architecture and has 2,048 CUDA cores, 64 Tensor Cores, and 2 Deep Learning Accelerator (DLA) engines. It can deliver an astonishing 275 TOPS of raw AI performance for models that have been optimized for 8-bit integer math.
For current Jetson customers, Orin features the same pin-out and footprint as the Jetson AGX Xavier.
This is not your Father’s (or Mother’s) Jetson Dev Kit
The first time I reviewed a Jetson Dev Kit, it arrived as a board, a daughterboard, and some screws. I think I had to buy my own small fan, an appropriate power supply, and I 3D printed a cheesy enclosure. The Orin Dev kit is an ode to design. A machined metal enclosure, with internal fans, and a magnetically attached cover for a PCI-e slot. It looks cool, and can draw power either from a barrel connector or a USB-C port.