Raspberry Pi 5 or Jetson: the real dividing line in robot compute
The difference is not compute. It is whether your program has a neural network on its real-time path. Answer that and the decision makes itself.
- Raspberry Pi 5
- 4/8 GB, about USD 60–80, no CUDA
- Jetson Orin Nano Super
- USD 249 dev kit, CUDA + TensorRT
- OS
- Both run Ubuntu / Debian-family
- Power
- RPi5 about 5–12 W; Orin Nano Super about 7–25 W
- Ecosystem
- RPi5 community an order larger; Jetson has the vendor model zoo
- Roles
- RPi5 = system and bridge; Jetson = perception and inference
- The spec, and what it measures
- The TOPS figure on the spec sheet is marketing: between INT8 peak throughput and your model's actual inference speed stand quantisation, memory bandwidth and batch size. The only metric that matters is your model's measured frame rate on the candidate board. A quantised YOLO-class detector runs at single-digit to low-teens FPS on a Pi 5 and at tens of FPS on an Orin Nano Super. The gap is real but smaller than the TOPS figures suggest — because the bottleneck is usually memory bandwidth, which marketing does not print.
- Price tiers
- USD 60–80 (RPi5 4/8 GB) → about 120 (RPi5 16 GB) → 249 (Orin Nano Super dev kit) → 500+ (Orin NX and up). The trigger for moving up a tier is a requirement, not a budget: without real-time neural inference on board, the 249-dollar tier is idle silicon.
- How to pick
- Three questions: ① is there a neural network on the real-time path? Yes — Jetson. No — Raspberry Pi. ② Do you need the CUDA ecosystem (model zoo, TensorRT, train/deploy consistency)? ③ Do power and cooling fit your chassis (both need deliberate thermal design in small builds)? The Pi 5's correct role is 'system and bridge': ROS 2 graph, drivers, logs, web UI — with real-time control delegated to an MCU and heavy inference to the cloud or a separate accelerator.
- Known pitfalls
- ① Putting inference inside the control loop — the loop rate collapses; the fix is separating perception and control into layers; ② cooling: the Pi 5 throttles noticeably under sustained load, and fanless Jetson modes throttle too; ③ JetPack versions bind hard to Ubuntu and ROS 2 versions — check the version matrix before flashing, a wrong matrix costs a day; ④ SD cards die under log-heavy systems within months — SSD, and this applies to the Pi too.
- What we checked
- We verified: prices and specs against the vendors' own pages; the JetPack–ROS 2 version matrix binding (official support tables); and that 'measured frame rate is the only metric that matters' is the consistent conclusion of community benchmarks. We did not run the comparison ourselves — FPS figures quoted here are community-scale; measure your own model once it exists. The grade stays accordingly.
This site's call
Default to the Raspberry Pi 5. Cheap, huge ecosystem, cheap to replace, and it covers everything from a remote-controlled car to SLAM and navigation. Move to a Jetson only when real-time neural inference enters the picture — at which point you will also know which metric to measure. The reverse order (buy the Jetson first, decide later) is the most common beginner waste.
Requirement → choice
| Requirement | Pick | Why |
|---|---|---|
| Teleoperation / data collection | Pi 5 (or less) | No real-time inference; a Jetson would be idle silicon |
| SLAM + navigation | Pi 5 | slam_toolbox and nav2 run comfortably on a Pi 5 |
| Real-time vision tracking / VLA inference | Jetson Orin Nano Super | CUDA ecosystem and TensorRT quantisation |
| Training policies | Neither | Training happens on a desktop GPU or in the cloud; the robot only infers |
Sources
- Raspberry Pi and NVIDIA official pages (prices, specs, power)
- JetPack and ROS 2 version support matrix
Objects in this entry
Last checked 2026-09-28