8-Camera Link Budget

Worked example

8-Camera Link Budget

Data-rate math for an eight-camera system of 2 MP ADAS sensors, from pixels to link utilization on FPD-Link and GMSL, through the deserializer hubs and into the SoC.

Sensor2 MP, 1920 × 1080 active (IMX390-class 1936 × 1096 is ~2% more)
OutputRAW12 over CSI-2, 30 fps (60 fps for the front camera)
Per camera0.78 Gb/s at 30 fps · 1.56 Gb/s at 60 fps
8 cameras6.2 Gb/s · 498 Mpx/s at 30 fps

RGB vs RCCC: same data rate

An automotive sensor puts one color filter on each photosite. RGGB (Bayer) gives color; RCCC replaces three of four filters with clear, giving about twice the sensitivity for night driving at the cost of color (only red is kept, for tail lights and signs). RCCB sits between. In every case the sensor outputs one raw sample per pixel, so the data rate depends only on resolution, bit depth and frame rate, not on the filter pattern.

The bit depth is where the choice matters. ADAS sensors capture 20–24 bits of HDR dynamic range (several exposures or a split pixel), then compand it on-chip to 12 bits (a piecewise-linear curve the ISP inverts). RAW12 is the common output. Sending uncompanded RAW20 or a processed YUV/RGB output costs more link bandwidth.

The math, step by step (RAW12, 30 fps)

  1. Active pixel rate. 1920 × 1080 = 2,073,600 pixels per frame; × 30 fps = 62.2 Mpx/s.
  2. Active bit rate. 62.2 Mpx/s × 12 bits = 746 Mb/s. That is the image itself.
  3. Vertical blanking. The sensor reads 1125 line periods per frame but only 1080 carry pixels. The link has to keep up during the active lines, so the rate it must sustain is 746 × 1125 / 1080 = 778 Mb/s. (Horizontal blanking does not add bits; the serializer's line buffer smooths it.)
  4. CSI-2 packet overhead. Each line is one long packet: 4-byte header + 2,880 bytes of RAW12 + 2-byte CRC, plus frame start/end short packets. 6 / 2880 = 0.2%. A few embedded-data lines (sensor metadata) add another ~0.2%. Result: ≈ 0.78 Gb/s per camera.
  5. Compare with the link's usable video rate, not its line rate. Coding and link-layer packets take roughly 20%: FPD-Link IV runs 7.55 Gb/s on the wire for 6 Gb/s of video. 0.78 / 4.8 ≈ 16% of a GMSL2 6G link; 0.78 / 3.33 ≈ 23% of an FPD-Link III link.
  6. Eight cameras. 8 × 0.78 = 6.2 Gb/s into the compute board, 498 Mpx/s for the ISP.
R_link = W × H × bpp × fps × (H_total / H) × (1 + overhead)= 1920 × 1080 × 12 × 30 × 1.042 × 1.002 ≈ 0.78 Gb/s

Every format, every link

Utilization is the camera's need divided by the link's usable video rate. Amber is above 80% (little margin for embedded data, retransmitted control traffic or a later sensor change); red does not fit.

Formatbits/pxActive Gb/sLink need Gb/s8 cams Gb/sFPD-Link III (UB953)GMSL2 3 Gb/sGMSL2 6 Gb/sFPD-Link IV (UB971)GMSL3 12 Gb/s
RAW10, 30 fps100.620.655.220%27%14%11%7%
RAW10, 60 fps101.241.3010.439%54%27%22%14%
RAW12 (RCCC or RGGB), 30 fps120.750.786.223%32%16%13%8%
RAW12 (RCCC or RGGB), 60 fps121.491.5612.547%65%32%26%16%
RAW20 HDR, uncompanded, 30 fps201.241.3010.439%54%27%22%14%
RAW20 HDR, uncompanded, 60 fps202.492.6020.878%108% ✕54%43%27%
YUV422 8-bit, 30 fps161.001.048.331%43%22%17%11%
YUV422 8-bit, 60 fps161.992.0816.662%87%43%35%22%
RGB888, 30 fps241.491.5612.547%65%32%26%16%
RGB888, 60 fps242.993.1124.994%130% ✕65%52%32%

Reading the table: a 2 MP camera never stresses its own link. Even RGB888 at 60 fps fits on GMSL2 6G or FPD-Link IV. The places that run out first are the 3 Gb/s GMSL2 mode and FPD-Link III with uncompressed HDR or RGB at 60 fps, and, for every format, the aggregation point on the compute board.

Eight cameras on one compute board

Cam 1Cam 2Cam 3Cam 40.78 Gb/s eachQuad hub 1UB960 / MAX96712 classCSI-2 ×4: 3.1 Gb/sCam 5Cam 6Cam 7Cam 80.78 Gb/s eachQuad hub 2UB960 / MAX96712 classCSI-2 ×4: 3.1 Gb/sSoC2 CSI-2 ports, ISP6.2 Gb/s · 498 Mpx/s totalRAW12, 30 fps, 8 cameras
The usual 8-camera layout: two quad hubs, each merging four links by virtual channel onto one 4-lane CSI-2 port. Numbers are for 2 MP RAW12 at 30 fps.
Case (8 × 2 MP)TotalPer quad hub (4 cams)Per CSI-2 lane, 4 lanesVerdict
RAW12, all 30 fps6.2 Gb/s3.1 Gb/s0.78 Gb/sComfortable on 1.5 Gb/s lanes (52%)
RAW12, front 60 fps, 7 at 307.0 Gb/s3.9 / 3.1 Gb/s0.97 Gb/sComfortable
RAW12, all 60 fps12.5 Gb/s6.2 Gb/s1.56 Gb/sNeeds 2.5 Gb/s lanes (62%) or two output ports per hub
YUV422, all 30 fps8.3 Gb/s4.2 Gb/s1.04 Gb/sOK on 1.5 Gb/s lanes (69%)
RAW20 HDR, all 30 fps10.4 Gb/s5.2 Gb/s1.30 Gb/sTight on 1.5 Gb/s lanes (87%); use 2.5 Gb/s or split ports

Check the hub's CSI-2 transmitter lane-rate limit and the SoC's receiver limit (each is commonly 1.5–2.5 Gb/s per D-PHY lane), and how many CSI-2 ports and virtual channels the SoC can accept. With the front camera at 60 fps on one hub, balance the other hub so neither port runs hot.

Bursts: why the serializer needs a line buffer

The sensor does not send at 0.78 Gb/s; it sends each line in a burst at its CSI-2 lane rate, then idles in blanking. At 4 lanes × 800 Mb/s = 3.2 Gb/s, a 2,880-byte line takes 7.2 µs out of a 29.6 µs line period (30 fps, 1125 lines). If the burst rate exceeds the link's usable rate, the serializer stores the difference: on a 2.4 Gb/s GMSL2 3G link, (3.2 − 2.4) Gb/s × 7.2 µs ≈ 5.8 kbit, well inside a line buffer. The rule: the link must carry the average rate within each line period; buffers absorb the burst inside it. Set the sensor's CSI-2 rate no higher than the serializer's CSI-2 input limit.

Beyond the cable: SoC and storage

Load (8 × 2 MP RAW12, 30 fps)MathResult
ISP pixel throughput8 × 62.2 Mpx/s498 Mpx/s; compare with the SoC ISP's rated Mpx/s
DDR write of raw frames (16-bit containers)498 Mpx/s × 2 B~1.0 GB/s
ISP read raw + write YUV420 (1.5 B/px)1.0 + 0.75 GB/s~1.75 GB/s more
Raw data logging6.0 Gb/s active ÷ 8 × 3600~2.7 TB per hour

Memory traffic for cameras is a few GB/s, small next to a Thor-class SoC's 273 GB/s but significant on a smaller SoC that also runs perception. Logging is the real cost for a data-collection robot: 2.7 TB per hour pushes toward on-sensor or on-board compression for fleet logging.

Calculator

Change the sensor or link and the numbers update. Usable rates for GMSL are an assumed ~80% of line rate; FPD-Link III uses the UB953's ~3.33 Gb/s CSI-2 input limit; FPD-Link IV uses TI's 6 Gb/s video figure.

The one-minute answer

If asked "what does 8 cameras need?"

A 2 MP ADAS camera, RCCC or Bayer, outputs one 12-bit companded raw sample per pixel: 1920 × 1080 × 12 × 30 is 746 Mb/s, about 0.78 Gb/s with vertical blanking and CSI-2 overhead. That uses under a quarter of one FPD-Link III link and about 16% of GMSL2 6G, so the per-camera link is never the issue. Eight of them are 6.2 Gb/s and about 500 Mpx/s. Two quad hubs each put 3.1 Gb/s onto a 4-lane CSI-2 port, under 1 Gb/s per lane. The pinch points are the hub and SoC CSI-2 lane rates at 60 fps or HDR-uncompanded formats, ISP throughput, DDR traffic, and about 2.7 TB per hour if you log raw.