Jetson Orin NX 16 GB
The step up for humanoids and mobile manipulators running multiple camera streams plus a policy network. 100 TOPS and 16 GB of shared memory.
The role it plays
Every serious robot needs two tiers. A microcontroller owns the control loop at a fixed rate with hard deadlines, because a missed deadline is a visible jolt. A Linux or GPU board owns perception, planning and networking, because those are bursty and best-effort. Nothing balance-critical belongs on the Linux side.
Typical use: Humanoids, mobile manipulators, multi-camera VLA inference.
What the numbers actually mean
The headline figure is 100 TOPS (INT8, sparse) (ai performance), with 1024-core Ampere as the next most load-bearing number (gpu). At $499 in the compute tier it sits as a capital purchase. At this price you are buying documentation, repeatability and support as much as metal, and the integration cost is usually larger than the part.
How it compares
Same subsystem, closest alternatives. The primary spec column is usually the one that decides it.
| Part | Price | Primary spec | Second spec |
|---|---|---|---|
| Jetson Orin NX 16 GBEdge GPU | $499 | 100 TOPS (INT8, sparse) | 1024-core Ampere |
| Raspberry Pi 5 (8 GB)SBC, high-level brain | $80 | 4× Cortex-A76 @ 2.4 GHz | 8 GB LPDDR4X |
| Jetson Orin Nano SuperEdge GPU | $249 | 67 TOPS (INT8, sparse) | 1024-core Ampere |
| ESP32-S3Real-time MCU + Wi-Fi | $5 | 2× Xtensa LX7 @ 240 MHz | 512 KB SRAM + 8 MB PSRAM typical |
| Teensy 4.1High-performance MCU | $32 | Cortex-M7 @ 600 MHz, hardware FPU | 1 MB + 8 MB PSRAM (optional) |
Where it shows up
Platforms in this atlas that use something in this class: Unitree G1, Boston Dynamics Spot, ANYmal D.