Low FPS issue with OAK-FFC-4P PoE (IMX378) on Jetson AGX

mohamed

Hello everyone,

I’m facing a severe FPS limitation when using an OAK-FFC-4P PoE with 4× IMX378-W sensors, and I’d appreciate guidance to understand where the bottleneck is.

1️⃣ Hardware setup

Device: OAK-FFC-4P PoE
Sensors: 4 × IMX378-W
Connection: PoE
Switch: TP-Link TL-SG1005LP
Host: NVIDIA Jetson AGX

2️⃣ Software environment

OS: Ubuntu (Jetson)
DepthAI: 2.30.0.0
Language: Python
OpenCV: via pip (standard)

3️⃣ Pipeline description

4 × ColorCamera
Each camera:
- Resolution: 1080p
- FPS requested: 30
- video output stream (full resolution)
- Image quality tuned (low denoise, auto WB)
All streams sent via XLinkOut and displayed with OpenCV

Simplified pipeline code

$$
$$

$$
cam = pipeline.create(dai.node.ColorCamera)
cam.setBoardSocket(socket)
cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
cam.setFps(30)
cam.setInterleaved(False)
cam.setColorOrder(dai.ColorCameraProperties.ColorOrder.BGR)

xout = pipeline.create(dai.node.XLinkOut)
xout.setStreamName(f"{name}_video")
cam.video.link(xout.input)
$$

Each stream is consumed with:

$$
$$

$$
device.getOutputQueue(stream, maxSize=4, blocking=False)
$$

FPS is measured per camera on the host.

4️⃣ Observed issue

Single camera: ~6 FPS
Four cameras: ~2 FPS per camera
This happens even without any neural network, just raw video streaming.
Lowering resolution does not reach 20 FPS.
CPU usage on the Jetson is not maxed out.

5️⃣ Diagnostics already performed

Bandwidth test

$$
$$

$$
python3 oak_bandwidth_test.py
$$

Results:

Downlink: 201.0 Mbps
Uplink: 253.9 Mbps

PoE / link test

$$
$$

$$
python3 poe_test.py
$$

Results:

Link speed: 1000 Mbps
Full duplex: OK
Boot mode: OK

So Ethernet bandwidth and PoE link appear healthy.

I also tested official DepthAI examples, and I still observe similarly low FPS.

6️⃣ Expected behavior

At least 20–30 FPS per camera
Full use of all 4 sensors
High image quality
Real-time preview

7️⃣ Questions / what I’m trying to understand

Is there a known limitation with:
- 4× IMX378
- PoE
- Jetson AGX
- Python + OpenCV display?
Is the bottleneck:
- On the device (ISP / encoder)?
- On XLink over Ethernet?
- On the host-side OpenCV display?
Would using:
- preview instead of video
- MJPEG / H.264 encoding
- Non-BGR format
- Larger queue sizes
  significantly improve FPS?

Goal

Maximum FPS
High image quality
Real-time multi-camera streaming

Any insight, recommended pipeline changes, or reference designs for 4× 1080p cameras on OAK-FFC-4P PoE would be greatly appreciated.

Thanks in advance.

OskarSonc

You’re basically hitting a hard bandwidth wall: cam.video is uncompressed, so 1080p 30 is already 0.7–1.4 Gbps per camera. With 4 cameras that’s multiple Gbps, which can’t fit over 1GbE so 6 FPS for 1 cam and FPS for 4 cams is expected even without any NN.

Also, your oak_bandwidth_test.py result 200 Mbps downlink is much lower than what you’d normally expect on a healthy 1GbE path, so there may be an additional network/host bottleneck (switch/NIC/MTU/offloads/cable), but even with full 900 Mbps you still won’t get 4× raw 1080p.

To get real-time FPS over PoE, use on-device encoding (H.264/H.265 or MJPEG) and stream the bitstream, or use preview/ISP-scaled output for live view. Avoid forcing BGR on the wire; only convert for display.

Thanks,
Oskar