Hi! I'm trying to stream both OV9282 mono cameras at 1280×800 simultaneously, as fast as possible (ideally the full 129 fps each) on an OAK-D-W, depthai v3.7.1, USB3 SuperSpeed. No depth, no NN, no encoding, just raw frames to the host. Single camera at 1280×800 hits \~129 fps fine. The moment I run both, each drops to \~75 fps, and total throughput barely rises above the single-camera rate. Things I've already tried (numbers are per-camera, 1280×800 stereo): Baseline (small pools, blocking queues) : \~35fps setRawNumFramesPool/setIspNumFramesPool = 16 + non-blocking queues : \~62fps \+ pipeline.setXLinkChunkSize(0) : \~75fps \+ manual exposure & AE/AWB lock : \~81fps This looks XLink-bound to me: aggregate transfer plateaus at \~154 MB/s regardless of stream count or host reader threads, and it tracks setXLinkChunkSize (65535 → \~108 MB/s, 0 → \~154 MB/s). We're only at \~1.2 Gbps on a 5 Gbps link, so it seems protocol-bound rather than PHY-bound. Dead ends I found: * On-device encoding doesn't help, the encoder needs NV12, and the ISP's NV12 output itself caps at \~45 fps/cam (23 stereo), below uncompressed RAW8. GRAY8/YUV420p are rejected by the encoder; YUV400p works but is slower (\~22). * cam.raw (PACK10) is slower (\~64 fps/cam) and clamps to 1280×720. * requestIspOutput() gives \~84 fps but at 1280×720 (same Mpix/s). * More XLink streams / per-stream reader threads, no change. Questions: * Is \~150–154 MB/s the expected XLink throughput ceiling for an RVC2 device over USB3 SuperSpeed, or should I be able to push it higher (firmware setting, setXLinkRateLimit, USB 3.2 Gen2 host, anything)? * Is there any way to feed the on-device VideoEncoder a luma-only / GRAY8 input so I can compress without paying the NV12 ISP cost? (My real goal is high-fps full-res stereo; compressed is acceptable if quality is reasonable.) * Is full-res 1280×800 stereo at \~120 fps simply not feasible on RVC2, and does RVC4 (OAK-4) remove this limit? Thanks!

Full-res (1280×800) stereo on OAK-D-W (RVC2) caps at ~75–81 fps/cam

Ivan

Hi!

I'm trying to stream both OV9282 mono cameras at 1280×800 simultaneously, as fast as possible (ideally the full 129 fps each) on an OAK-D-W, depthai v3.7.1, USB3 SuperSpeed. No depth, no NN, no encoding, just raw frames to the host.

Single camera at 1280×800 hits ~129 fps fine. The moment I run both, each drops to ~75 fps, and total throughput barely rises above the single-camera rate.

Things I've already tried (numbers are per-camera, 1280×800 stereo):

Baseline (small pools, blocking queues) : ~35fps
setRawNumFramesPool/setIspNumFramesPool = 16 + non-blocking queues : ~62fps
+ pipeline.setXLinkChunkSize(0) : ~75fps
+ manual exposure & AE/AWB lock : ~81fps

This looks XLink-bound to me: aggregate transfer plateaus at ~154 MB/s regardless of stream count or host reader threads, and it tracks setXLinkChunkSize (65535 → ~108 MB/s, 0 → ~154 MB/s). We're only at ~1.2 Gbps on a 5 Gbps link, so it seems protocol-bound rather than PHY-bound.

Dead ends I found:

On-device encoding doesn't help, the encoder needs NV12, and the ISP's NV12 output itself caps at ~45 fps/cam (23 stereo), below uncompressed RAW8. GRAY8/YUV420p are rejected by the encoder; YUV400p works but is slower (~22).
cam.raw (PACK10) is slower (~64 fps/cam) and clamps to 1280×720.
requestIspOutput() gives ~84 fps but at 1280×720 (same Mpix/s).
More XLink streams / per-stream reader threads, no change.

Questions:

Is ~150–154 MB/s the expected XLink throughput ceiling for an RVC2 device over USB3 SuperSpeed, or should I be able to push it higher (firmware setting, setXLinkRateLimit, USB 3.2 Gen2 host, anything)?
Is there any way to feed the on-device VideoEncoder a luma-only / GRAY8 input so I can compress without paying the NV12 ISP cost? (My real goal is high-fps full-res stereo; compressed is acceptable if quality is reasonable.)
Is full-res 1280×800 stereo at ~120 fps simply not feasible on RVC2, and does RVC4 (OAK-4) remove this limit?

Thanks!

Oskar-Luxonis

Hey Ivan - measurements look like an RVC2 host-transfer limit.

1280x800 RAW8 is 1.024 MB/frame. Two cameras at 129 FPS would need 264 MB/s before overhead. Your 154 MB/s plateau corresponds to:

154 / (1.024 * 2) ~= 75 FPS/camera
So the 75 FPS/cam result is expected for this uncompressed dual-stream workload. setXLinkChunkSize(0) is the right optimization, and your numbers show it helps. setXLinkRateLimit() will not help; it only throttles.

For encoding: VideoEncoder supports NV12, and GRAY8 is converted to NV12 internally, so there is no true luma-only encoder path that avoids the NV12/ISP cost.

So on OAK-D-W/RVC2, full 1280x800 stereo at 120 FPS to host uncompressed is not really feasible over the normal USB/XLink path.

For OAK4 the OV9282 left/right mono cameras are max 60 FPS. The center IMX586 camera has HFR modes, e.g. 1280x720 up to 480 FPS and 1920x1080 up to 240 FPS, but that does not apply to the stereo OV9282 pair.

Best,
Oskar

Ivan

Thanks for the detailed breakdown, Oskar, that makes sense.

Is there any way to get higher fps/cam while keeping the resolution at 1280×800? Any configuration or settings you'd suggest I try?

We're only at ~1.2 of 5 Gbps, so it looks protocol-bound. Is 154 MB/s a hard firmware limit, or is there any knob, USB bulk transfer size, in-flight XLink packet/buffer count, bootloader/firmware setting, USB 3.2 Gen2 host — that pushes it higher? And is the ceiling per-link or per-device: if I split left/right across two separate USB endpoints/XLink connections, does aggregate throughput scale, or is there one shared device-side queue?

Oskar-Luxonis

Ivan - I don’t know of any user-facing setting that pushes this higher on RVC2.

With the same OV9282 setup, my best result was also pipeline.setXLinkChunkSize(0): 152-155 MB/s aggregate, or 75 FPS/camera at 1280x800 RAW8/GRAY8. Default chunking was lower.

setXLinkRateLimit() only throttles. A USB 3.2 Gen2 host won’t help, since the RVC2 path is still USB3/XLink. Splitting into more queues/threads/streams also won’t scale aggregate throughput; the streams share the same device-side USB/XLink path.

So for one RVC2 device at 1280x800:

single mono: 129 FPS
dual mono: 75 FPS/camera
compression: slower here, because NV12/YUV400/encoder becomes the bottleneck

I’d treat 150-155 MB/s as the practical RVC2 ceiling for this workload.