Jetson

Jetson#

NVIDIA’s embedded AI compute platform. ARM CPU paired with a proper CUDA-capable GPU and dedicated tensor / DLA accelerators on the same die; runs vision, ASR, and LLM inference on-device with no cloud round-trip. The operator’s offline ML host on target.

Lineup#

Module

CPU

GPU / NPU

RAM

AI perf

Note

Jetson Orin Nano

6× Cortex-A78AE

1024-core Ampere + 32 tensor

4 / 8 GB

up to 67 TOPS

Entry-level Orin. The current default for offline CV / ASR work.

Jetson Orin NX

6 or 8× Cortex-A78AE

1024-core Ampere + 32 tensor

8 / 16 GB

up to 157 TOPS

Step up; more RAM, higher clocks. SO-DIMM form factor.

Jetson AGX Orin

8 or 12× Cortex-A78AE

2048-core Ampere + 64 tensor + 2× DLA

32 / 64 GB

up to 275 TOPS

The big one. 100 × 87 mm; integrates a 40 GbE-capable PHY.

Jetson Xavier NX

6× Carmel

384-core Volta + 48 tensor

8 / 16 GB

21 TOPS

Older generation; still widely deployed and supported.

Jetson Nano (legacy)

4× Cortex-A57

128-core Maxwell

2 / 4 GB

0.5 TFLOPS

End-of-life as of 2024 but still in the wild; useful for cheap CV experiments.

I/O#

Interface

Notes

40-pin GPIO header (devkit)

Mostly Pi-compatible pinout; runs at 3.3 V logic.

PCIe / M.2

One or more M.2 slots for NVMe SSDs and Wi-Fi modules; Orin and AGX add PCIe Gen 4 lanes.

USB

Multiple USB 3.x ports plus a Type-C device port for flashing.

Network

Gigabit Ethernet on Orin Nano; multi-gig and 10 GbE on Orin NX and AGX.

Display

DisplayPort and HDMI on the devkits.

Camera

Two or four MIPI CSI-2 lanes; pairs with Sony IMX and OnSemi sensors.

Operator use#

  • Offline computer vision. YOLOv8, MediaPipe, segmentation pipelines, OCR; the Jetson runs them at frame rates that matter, with no upstream traffic.

  • Offline ASR / speech-to-text. Whisper (small / base / medium) runs comfortably on Orin Nano.

  • On-device LLM inference. Llama 3 8B Q4 fits on Orin NX; larger models need AGX.

  • Drone payload. AGX Orin is the de facto autonomy / vision computer on serious UAVs.

  • Edge sensor fusion. Lidar, camera, IMU, and radar all decoded on-device.

Tools#

Tool

Effect

JetPack

NVIDIA’s BSP and SDK; bundles L4T (Linux for Tegra), CUDA, cuDNN, TensorRT.

SDK Manager

Flash a developer kit from a host Linux machine.

tegrastats

Realtime CPU / GPU / RAM utilization in the shell.

nvpmodel

Set power modes (e.g. MAXN for full performance).

jetson-containers

Prebuilt containers for PyTorch, TensorRT, transformers.

DeepStream

NVIDIA’s reference video-analytics pipeline framework.

References#

Contents