[BUG] OAK-FFC-4P-POE (RVC2): 4th camera (CAM_D/socket 3) yields 0 frames streaming 4x AR0234

[siyuanfeng-lbm]

Summary

On an OAK-FFC-4P-POE (RVC2 / Myriad X) with four OAK-FFC AR0234 (2-lane) modules — one in each socket (CAM_A, CAM_B, CAM_C, CAM_D) — all four cameras enumerate, but when streaming all four simultaneously the 4th camera (CAM_D / socket 3) yields 0 frames, while CAM_A/B/C stream normally. Any 3 of the 4 sockets work (including both 4-lane sockets A+D together); only the 4th, whichever is allocated last (CAM_D), fails.

Hardware

• Board: OAK-FFC-4P-POE — EEPROM productName=OAK-FFC-4P-POE, boardName=NG2093, boardRev=R4M1E4, MXID 14442C10414E91D700, bootloader 0.0.28, platform RVC2.
• 4× OAK-FFC AR0234 (2-lane, 1920x1200), one per socket A/B/C/D.
• PoE, point-to-point (device at link-local 169.254.1.222).

Goal

Stream all 4 cameras simultaneously (MJPEG), full resolution 1920x1200, ~20 fps each.

Symptom

• getConnectedCameraFeatures() returns all four AR0234.
• Pipeline with one Camera→VideoEncoder(MJPEG) per socket: CAM_D delivers 0 frames; CAM_A/B/C deliver 20 fps.
• Dropping any one camera → the remaining 3 all stream fine (tested A+C+D, D+A+B — both include the two 4-lane sockets).

Reproduced identically across

• Python (depthai v3.7.1, release device firmware 8d6a04d380ce182e0cb3a4694309426370295a9f) — single-consumer and thread-per-camera pollers.
• C++ built from develop (device firmware 45274630e27b04617985d3e80360f15052a99686).
• With and without BoardConfig.mipi4LaneRgb = false (no effect; it also appears unused in host code).
• Power-cycling the board does not change it.

What this rules out

• Not lane bandwidth: AR0234 modules are 2-lane → 4×2 = 8 lanes total.
• Not I2C: all four enumerate (unlike the shared-I2C FFC-ToF-33D case in [Bug] Unable to access to 4 TOFs from 4 S5K33D through OAK-FFC-4P depthai-python#1007).
• Not firmware version: reproduced on both release and develop firmware.

Related but not identical: #857 (CAM_D not detected — fixed) and luxonis/depthai-python#1007 (ToF, shared-I2C-limited to 3).

Minimal reproducer (C++)

Built against depthai-core (BUILD_SHARED_LIBS=ON, DEPTHAI_OPENCV_SUPPORT=ON). Prints per-socket frame counts over 12 s; CAM_D shows 0.

#include <atomic>
#include <chrono>
#include <cstdio>
#include <map>
#include <memory>
#include <string>
#include <thread>
#include <vector>
#include "depthai/depthai.hpp"

int main() {
  auto dev = std::make_shared<dai::Device>(dai::DeviceInfo("169.254.1.222"));
  std::printf("connected %s fw-driven test\n", dev->getDeviceName().c_str());
  std::vector<std::pair<std::string, dai::CameraBoardSocket>> socks = {
      {"CAM_A", dai::CameraBoardSocket::CAM_A},
      {"CAM_B", dai::CameraBoardSocket::CAM_B},
      {"CAM_C", dai::CameraBoardSocket::CAM_C},
      {"CAM_D", dai::CameraBoardSocket::CAM_D}};
  dai::Pipeline p(dev);
  std::map<std::string, std::shared_ptr<dai::MessageQueue>> qs;
  for (auto& [name, s] : socks) {
    auto cam = p.create<dai::node::Camera>()->build(s);
    auto* out = cam->requestOutput({1920, 1200}, dai::ImgFrame::Type::NV12,
                                   dai::ImgResizeMode::CROP, 20.0f);
    auto enc = p.create<dai::node::VideoEncoder>();
    enc->setDefaultProfilePreset(20.0f,
                                 dai::VideoEncoderProperties::Profile::MJPEG);
    enc->setQuality(80);
    out->link(enc->input);
    qs[name] = enc->out.createOutputQueue();
  }
  p.start();
  std::atomic<bool> stop{false};
  std::map<std::string, std::atomic<int>> counts;
  for (auto& [name, s] : socks) counts[name] = 0;
  std::vector<std::thread> ths;
  for (auto& [name, q] : qs) {
    ths.emplace_back([&, name, q] {
      while (!stop.load()) {
        bool to = false;
        auto f = q->get<dai::EncodedFrame>(std::chrono::milliseconds(200), to);
        if (!to && f) counts[name]++;
      }
    });
  }
  std::this_thread::sleep_for(std::chrono::seconds(12));
  stop = true;
  for (auto& t : ths) t.join();
  std::printf("=== 4 cams @20 q80, develop firmware ===\n");
  for (auto& [name, s] : socks)
    std::printf("  %s: %d frames (%.1f fps)\n", name.c_str(),
                counts[name].load(), counts[name].load() / 12.0);
  return 0;
}

Example output (device IP hard-coded above):

connected OAK-FFC-4P-POE fw-driven test
=== 4 cams @20 q80, develop firmware ===
  CAM_A: 241 frames (20.1 fps)
  CAM_B: 235 frames (19.6 fps)
  CAM_C: 235 frames (19.6 fps)
  CAM_D: 0 frames (0.0 fps)

Questions

1. Is streaming 4 cameras simultaneously supported on the OAK-FFC-4P (RVC2), or is there a hard limit on the number of simultaneous MIPI camera streams on the Myriad X for this board?
2. If supported, what is the correct way to enable it (BoardConfig, flashed board/calibration config, GPIO, or a firmware build)? Is there a way to force the 4-lane sockets (A/D) to operate in 2-lane mode so all four fit?
3. If it is a hardware limit, which board supports 4 concurrent 2-lane cameras (OAK-FFC-6P, or an RVC4-based board)?

Thanks!

[aljazdu]

Hi @siyuanfeng-lbm , thanks for a detailed report you've shared with us!

We've been able to reproduce the issue and are working on an update, which should ideally be added in the next release.

In the meantime, could you please confirm whether you have the same issues when using the release version v3.6.1 or below - as it seems there might be an issue which was introduced in the v3.7.1.

Regarding your questions:

1. Yes it is supported and will work after an update from our side.
2. There is no additional work needed to enable this, and should work out of the box.
3. OAK4 FFC will support this and be much more stable. RVC2 does support 4 concurrent streams but due to low overall compute, the device is quite stressed out.

Thanks again!

[siyuanfeng-lbm]

Thanks for the reply 3.6.1 seems to stream all four. Do you have a timeline for the fix?

[aljazdu]

@siyuanfeng-lbm the new release which fixes the issue will be released end of next week at the latest