Edge AI Platform Architecture
Notes on the edge AI platform I’m building for robotics work.
Compute: Jetson Orin Nano Super
NVIDIA’s Jetson line is the standard for edge AI that needs real GPU compute. The Orin Nano Super hits the right balance:
- 67 TOPS INT8 inference
- 1024 CUDA cores
- Hardware video encode/decode
- Full Linux environment
- Solid ROS2 support
Powerful enough for real-time inference on multiple camera streams. Good ecosystem for robotics.
Why Not Raspberry Pi
The Pi 5 works for many things, but ML inference isn’t one of them. No GPU, limited memory bandwidth. For anything beyond basic classification, you need actual compute.
Stack
- JetPack 6 / Ubuntu 22.04
- CUDA 12, cuDNN, TensorRT
- PyTorch 2.1
- ROS2 Humble (for robotics integration)
- Custom inference pipeline
More on the perception stack and sensor fusion in future posts.