Does NVCSI on Jetson AGX Thor support CSI-2 lane rates above 1.5 Gbps/lane, and how should it be configured?

Hello,

We have 4× ISX031 cameras (1920×1536 YUV422 @ 30 fps) behind a MAX96712 deserializer, aggregated onto a single 4-lane D-PHY CSI-2 output on CSI Port A. At 1.5 Gbps/lane the link is ~94% utilised, so we want to raise the CSI-2 output rate.

Setup: Jetson AGX Thor Developer Kit, L4T R39.2.1, kernel 6.8.12-1021-rt-tegra. The deserializer driver is our own out-of-tree driver.

configured

MAX96712 CSI/PHY registers (rate encoding from the in-tree staging max96712.c: value = (Mbps / 100) | BIT(5)):

0x1D00 = 0xF4 / 0x1E00 = 0xF4      // assert PHY resets
0x0415 = 0x0418 = 0x041B = 0x041E = <RATE>   // 0x2F=1.5G, 0x34=2.0G, 0x39=2.5G
0x1D00 = 0xF5 / 0x1E00 = 0xF5      // release PHY resets
0x08A2 = 0x34 ; 0x040B = 0x02 ; 0x08A0 = 0x84

Device tree, per sensor mode: pix_clk_hz = serdes_pix_clk_hz = 375000000 (1.5 G), 500000000 (2.0 G), 625000000 (2.5 G).

Observed

Verified on the board: register readback matches, all 4 GMSL links locked, and the RCE log confirms NVCSI is programmed to match — e.g. at 2.5 G: Lanes: 4, MIPI clock rate: 1250000 kHz, Physical rate: 2500000 Kbps.

MAX96712 rate DT pix_clk_hz Result
1.5 Gbps (0x2F) 375000000 Frames arrive but 100% are discarded: corr_err: discarding frame N, flags: 0, err_data 131072
2.0 Gbps (0x34) 500000000 No frames: uncorr_err: request timed out after 2500 ms
2.5 Gbps (0x39) 625000000 No frames: same uncorr_err timeout

Questions

  1. Does NVCSI on Jetson AGX Thor support CSI-2 lane rates above 1.5 Gbps/lane, and what is the maximum supported rate per lane?
  2. If it is supported, what configuration is required? We already match the device tree pix_clk_hz / serdes_pix_clk_hz to the target rate and NVCSI reports the corresponding Physical rate. Is anything else needed — additional device tree properties, D-PHY tuning, deskew, settle time, or calibration?
  3. How should corr_err / err_data = 131072 (0x20000) be decoded? It appears on every frame at 1.5 Gbps. Is there a documented bit mapping, or a debugfs/sysfs interface exposing the underlying NVCSI error status?

We understand our deserializer driver is custom and out of scope for NVIDIA support — we are asking specifically about the Tegra/NVCSI/device-tree side.

log

# 1.5 Gbps
[3198.222526] tegra-camrtc-capture-vi tegra-capture-vi: corr_err: discarding frame 1, flags: 0, err_data 131072
[3198.255831] tegra-camrtc-capture-vi tegra-capture-vi: corr_err: discarding frame 2, flags: 0, err_data 131072

# 2.5 Gbps — RCE stream configuration
[ 375.309653] [RCE] stream_id: PP 0, csi_port: PORT A
[ 375.309655] [RCE] Partition: CIL A, LP bypass: Enabled, Lanes: 4
[ 375.309657] [RCE] MIPI clock rate: 1250000 kHz
[ 375.309658] [RCE] Physical rate: 2500000 Kbps
[ 378.x]      tegra-camrtc-capture-vi tegra-capture-vi: uncorr_err: request timed out after 2500 ms

While the output data rate > 1.5G need send deskew word.

Skew calibration is required if sensor or deserializer is using DPHY, and the output data rate is > 1.5Gbps.
An initiation deskew signal should be sent by sensor or deserializer to perform the skew calibration. If the deskew signals is not sent, the receiver will stall, and the capture will time out.
You can calculate the output data rate with the following equation:

Output data rate = (sensor or deserializer pixel clock in hertz) * (bits per pixel) / (number of CSI lanes)

err_data 131072 = 0x00020000 = VI5_TD_STATUS_PIXEL_INCOMPLETE.

Meaning: VI saw the frame’s pixel payload as incomplete and dropped/reported the frame accordingly

Hi

Sony ISX031 → MAX96717 → MAX96712 → NVCSI, YUV422 16-bit, 1920×1536@30, 4 D-PHY lanes, destination CSI port A.

Configuration used:

/* MAX96717: CSI-2 RX deskew */
max96717_write_reg(dev, 0x0331, 0x70);

/* MAX96712: CSI TX pipe deskew */
max96712_write_reg(dev, 0x0903, 0x80);  // DESKEW_EN Pipe 0
max96712_write_reg(dev, 0x0943, 0x80);  // DESKEW_EN Pipe 1
max96712_write_reg(dev, 0x0983, 0x80);  // DESKEW_EN Pipe 2
max96712_write_reg(dev, 0x09C3, 0x80);  // DESKEW_EN Pipe 3
/* sensor mode0 */
pix_clk_hz = "625000000";
serdes_pix_clk_hz = "625000000";
deskew_initial_enable = "true";

MAX96712 also writes the PHY rate code (0x0418 etc.) — for 2.5 Gbps 0x39, and the alternate memory map 0x0933/0x0973/0x09B3/0x09F3 = 0x10.

Results (/dev/video0, one camera):

Lane rate DT clocks RCE Physical rate RCE MIPI clock rate corr_err uncorr_err Frames
1.5 G 375 MHz 1500000 Kbps 750 MHz 600 frames / 20 s, all discarded 0 0
2.0 G 500 MHz 2000000 Kbps 1000 MHz 0 timeout 2500 ms 0
2.5 G 625 MHz 2500000 Kbps 1250 MHz 0 timeout 2500 ms 0
  • 1.5 Gbps: frames arrive at 30 fps but every frame is discarded — corr_err: discarding frame N, flags: 0, err_data 131072.
  • 2.5 Gbps: identical result with all four pipe deskew registers and MAX96717 0x0331 enabled, with them disabled, and with the vendor PHY1-only sequence.
  • Present at all three rates: verify_rate_range: FMON_NVCSI: rate 5003892 below min 15000000, fmon_update_config: FMON_NVCSI: detected fault 0x80, CIL rate: 270000000 Hz.

Could you please advise where I should make changes to support CSI output above 1.5 Gbps — the GMSL driver configuration, the device tree, or another driver? Any guidance would be greatly appreciated. Thank you very much for your time and help.

Get the trace log for 1.5G for the detail then others rate.

Thanks

Hi @ShaneCCC

1.5Gbps

dmesg-1p5g.txt (82.4 KB)

trace-1p5g.txt (281.9 KB)

2.5Gbps

trace-2p5g.txt (21.9 KB)

dmesg-2p5g.txt (111.6 KB)

Could you please help review these logs and advise on the issue?

Thanks a lot!

Don’t see any error in the 1.5G trace.

In the kernel the internal error-recovery decode, data=0x00020000 is reported as PixelIncomplete, and the corresponding capture error is VI_SHORT_FRAME, meaning FE arrived before the normal number of pixels had appeared.

So the practical read is:

  • not a pure driver timeout by itself — this specific line says the incoming frame’s pixel payload was incomplete.
  • If it repeats, first check:
    • sensor/deserializer output is not truncating the frame,
    • active height / frame timing matches what VI expects,
    • embedded-line configuration is correct,
    • CSI source/link integrity is clean.

For the 2.5G trace log shows didn’t receive any data from the NVCSI/VI