sahil_ramani / field_v3.2
home writing projects resume about contact
ONLINE · PST
← /projects
SIDE PROJECT

Jetson Nano ML pipeline

Personal/2020 — 2021/Hobby project
PythonJetPackJupyterPyTorchC++

Personal exploration - blog series on setting up Python, Jupyter, and PyTorch on edge hardware.

The NVIDIA Jetson Nano is a $99 board that runs CUDA. It’s also a board that ships with a stack of opinions about Python versions, Jupyter, conda flavors, and PyTorch wheels - most of which were either out of date or didn’t quite agree with each other when I started poking at it.

This is the side-project I kept poking at for about a year: build a usable ML development environment on the Nano end-to-end, then write down everything I learned so other people didn’t have to find the same dead ends.

The posts

The series ran across most of 2020-2021. The most useful ones, roughly in the order you’d read them:

Why bother

The Nano is small, cheap, low-power, and runs the same CUDA you’d write for a workstation. That’s a useful constraint - it forces you to think about what your model actually needs to do, not what it’d be nice to throw GPU at. A lot of the lessons (memory pressure, kernel launch overhead, the cost of moving data) translate straight back to a server-class deployment.

The posts get steady traffic from people Googling specific error messages on Jetson hardware, which is exactly the audience I wrote them for.