hardware//NVIDIA//Jetson

Jetson is NVIDIA's microcomputer platform, built around the Orin SoC (Cortex-A cores and a hardwired Ampere GPU). It is both educational and industrial, and it is used in robotics, autonomous vehicles and edge ML.

Jetson is NVIDIA's microcomputer platform, built around the Orin SoC (Cortex-A cores and a hardwired Ampere GPU). It is both educational and industrial, and it is used in robotics, autonomous vehicles and edge ML.

In a robot or a drone it is the companion computer: drawing on the order of tens of watts, it runs CNNs and mid-sized transformers at tens of frames per second for perception (detection, segmentation, depth, visual odometry) and sends its conclusions to the flight controller or the motion controller, which keeps the fast loops. A detection network taking 10 to 30 ms there is fine for perception at 30 Hz and useless for a drone's attitude loop, whose deadline is 2 ms and whose worst case must be bounded (embedded system). Its power is also a flight-time cost: on a small quadcopter hovering on about 200 W, a 15 W board takes some 7 % of the endurance before its weight is counted.

Multiple Orins can sit on one PCB, connected through NVLink, for scaled-out edge compute, and CUDA enables deterministic programming of the GPU cores. Models trained on a data-centre GPU usually reach it through TensorRT, often quantized, and that converted artifact is the one to validate (training-serving skew).