Grab() takes a fixed faster path on ~1/3 of frames on two units, persists across restart and Camera::reboot (SDK 5.2.3, NEURAL)

Hi @Myzhar, Related background is in 11545, but this looks like a different behavior.

Two of our seven on-site boxes have switched into a state where about a third of grab() calls take a fixed, much shorter time than normal. The camera looks fine, depth looks fine by the one measure we have, and nothing we’ve tried clears it. So the real question is: what is grab() doing differently on those frames?

Setup

  • ZED Box Mini (Orin Nano 8GB, 15W mode), ZED 2i over USB, ZED SDK 5.2.3
  • HD1080 @ 30 fps, DEPTH_MODE::NEURAL, positional tracking + object detection enabled
  • One process per box, the same config on all seven

What we measured

Five samples a minute apart, each over the last 60 frames:

5 healthy boxes 2 affected boxes
Share of loop iterations that take the very next camera frame (33 ms image-timestamp gap) 0.00 - 0.02 0.27 - 0.45
grab() wall time, median ~60 ms 47 - 55 ms
grab() wall time, 10th percentile 46 - 59 ms 36.3 - 37.0 ms on BOTH boxes
Loop rate ~15 Hz ~19.5 Hz

On a healthy box every grab() takes ~60 ms and every iteration skips exactly one camera frame (67 ms gaps). On the two affected boxes there are two distinct grab() times, ~60 ms and a tight ~37 ms, and the ~37 ms value is the same on both units. That looks like a different code path, not noise.

What we ruled out

  • The camera. Image timestamps still advance by 33 ms steps, so it still delivers 30 fps. Exposure and gain read the same range as the healthy boxes (exposure 90 - 100 everywhere). Repeated frames are ~0, so these are new images.
  • Depth quality, roughly. The valid-pixel ratio over the whole frame stays at each box’s normal value.
  • The host config. Same power mode, GPU at max clock, same SDK version, and the optimized model files were built in September, before either box changed.
  • Time of day / lighting. The rate is flat from 05:00 to 22:00.

What clears it (nothing so far)

  • Process restart: no (daily, and dozens of times since)
  • Camera::reboot(sn, true): no (tried once on one of the boxes)
  • Full power cycle: not tried yet

Timeline

  • Box A switched on 2026-11-04, about an hour after a scheduled close + reopen of the camera at 05:30.
  • Box B came up in this state straight from the same 05:30 reopen on 2026-11-11.
  • No USB errors or re-enumerations at either onset. Both boxes did develop the worst USB link error counts in our fleet in the days AFTER, and both have since hit the zero-filled image fault from 11545 more than the others. That could be coincidence on two boxes, but I didn’t want to leave it out.

Questions

  1. Is there a path inside grab() that skips or reuses part of the NEURAL depth computation on some frames? ~37 ms vs ~60 ms is consistent with that, but I may be missing something.
  2. If so, what puts the SDK into it, and why would it survive a process restart and a Camera::reboot? Is there any state the SDK persists on disk or in the driver that we should clear?
  3. Is there a way to detect it from the API, or to force the normal path?
  4. Is this known, and is it different on 5.4.x or 5.5?

Happy to send per-box logs, the depth/left/right captures, or run a test build on one of the affected boxes. We can also try a full power cycle if that would tell you something.