anya_ml_eng
4 days ago
Third option nobody mentions enough in 2026: Rockchip RK3588-based boards — the Orange Pi 5 Plus and Radxa ROCK 5B both have an NPU at 6 TOPS, run full Debian, and cost roughly a quarter of what a comparable Jetson Orin NX setup runs. RKNN Toolkit 2 handles model deployment and has matured a lot in the past year. I've been running a pose estimation pipeline on one for a physiotherapy monitoring application at 30 FPS and it's been rock solid for six months. RKNN isn't as battle-tested as TensorRT and the community documentation still has gaps, but if your budget is tight and you don't need NVIDIA's ecosystem, benchmark it before committing to Jetson pricing.
coral_user_cw
5 days ago
Defending Coral for the right use case: for fixed-function IoT deployments where the model doesn't need frequent updates, the Dev Board Mini is genuinely excellent. We run a food freshness classifier on a conveyor belt — single MobileNetV2 model, quantized to int8, sitting at 7.2MB on the Edge TPU (just under the 8MB limit) — at 120 FPS with under 10ms latency, and the board never exceeds 3W. Replicating that power profile with Jetson hardware isn't possible. The tooling has also improved significantly; TFLite Model Maker now integrates Edge TPU compilation into the training workflow rather than it being a separate pain. Where Coral completely falls down is multi-model pipelines, dynamic model switching, or anything requiring a real Linux userspace — those are Jetson territory, full stop.
j_robotics
6 days ago
The DeepStream point is huge. I ran a 4-camera parking lot monitoring system on Jetson Orin NX and it handled four simultaneous 1080p30 streams with YOLOv8m detection at roughly 12W total board power — that's remarkable for the compute being delivered. Here's a minimal DeepStream pipeline config that got me from zero to working in about an hour once the SDK was installed. The key is getting your source type and sink type right before anything else:
# deepstream_app_config.txt — minimal 4-cam config
[application]
enable-perf-measurement=1
perf-measurement-interval-sec=5
[source0]
enable=1
type=4 # 4 = RTSP source
uri=rtsp://192.168.1.10:554/stream1
num-sources=1
[primary-gie]
enable=1
gpu-id=0
model-engine-file=yolov8m.engine # TRT engine, NOT ONNX
batch-size=4
network-type=0
interval=0
[sink0]
enable=1
type=3 # 3 = RTSP output
sync=0
rtsp-port=8554
Always export your model to a TensorRT engine directly on the target device with FP16 precision — never transfer an engine built on a different machine, TRT engines are hardware-specific and will silently fail or crash. The speedup versus ONNX Runtime is typically 3–5x on Orin hardware.
edge_vision_guy OP
1 week ago
I've deployed both in actual production environments over the past year so here's a real-world comparison, not a spec sheet one. The Jetson Orin NX 16GB at 10–15W is the better choice when you need flexibility: it runs full Linux, you can run multiple models concurrently via CUDA streams, and NVIDIA's DeepStream SDK handles multi-camera pipelines gracefully. The Google Coral Dev Board Mini — built around NXP i.MX 8M Mini plus the Edge TPU — is niche but genuinely brilliant for a single, fixed inference task. It boots in under 3 seconds and idles around 2W, which matters enormously in battery-backed outdoor installations. The Coral's hard limit is the 8MB model size cap on the Edge TPU co-processor — anything bigger spills to the ARM CPU and you lose most of the latency advantage. Your entire model pipeline has to be designed around that constraint before you write a single line of training code.
