Pixel
Light enters through the microlens and colour filter into a pinned photodiode. Charge sits on a pinned storage node and is read with true correlated double sampling — the low-noise global-shutter route at this pitch. Voltage-domain storage at 4 µm publishes around 3.8–4.3 e⁻, set by capacitor reset noise.
| Block | Choice | Why |
|---|---|---|
| Shutter | BSI charge-domain global shutter | True CDS. Voltage-domain pixels give up low-light colour. |
| Pitch | 4.0 µm, 4.5 µm option | 4.0 µm collects about 1.8× the light of 3.0 µm. 4.5 µm adds about 1.3×, on a larger die. |
| Shadows | Dual conversion gain | High gain for the floor. Low gain for a deep full well. Both, every frame. |
| Highlights | Lateral overflow, same exposure | Linear. No motion tear. Sized so colour survives the join. |
| Capacitor | In silicon, baseline | A bonded capacitor wafer only if the test chip requires it. |
| Optics | Full-depth backside isolation, 0° chief-ray microlens | The pixel accepts an f/1.4 cone. Front-side global shutters at this pitch lose about half of it. |
| Readout | 4 samples per row, 2 ADCs per column, merge on chip | 90 fps survives HDR. Calibration rides with the merge. |
| Output | 12-bit companded or 16-bit, 4-lane MIPI. Path and clip flags. | Fits current robot processors. A flagged pixel can be dropped. |
Charge-domain global shutter, dual gain, overflow storage and deep isolation each exist in the literature. Tohoku showed overflow-capacitor global shutter. Sony, OmniVision, Samsung and SmartSens have their own versions. The product is the measured result: colour held across one exposure, at 90 fps, with the cameras already in time.
Set aside for this generation: voltage-domain storage, a converter in every pixel, colour routers (gains shown below about 1.4 µm), organic film, photon counting, stacked colour.
Targets
Tier 1 is every block already shown somewhere, still to be shown together. Tier 2 is the stretch, open until the test chip measures overflow capacity and the join. Benchmark: SmartSens SC233HGS. n.p. = not published.
| Parameter | SC233HGS Published | Tier 1 Target | Tier 2 Target |
|---|---|---|---|
| Array | 1920 × 1200, 3.0 µm | 1920 × 1200, 4.0 µm | 4.5 µm option |
| Colour QE, G / B / R | n.p. | ≥65 / 50 / 55% | ≥72 / 60 / 63% |
| Read noise, high gain | 3.24 e⁻ | ≤2.0 e⁻ | ≤1.5 e⁻ |
| Linear full well | ~5.9 ke⁻ | ≥25 ke⁻ | ≥80 ke⁻ |
| Linear range at 90 fps | 62.65 dB | ≥80 dB | ≥90 dB |
| SNR at each HDR join | n.p. | ≥30 dB | ≥35 dB goal |
| Parasitic light | ~ −92 dB | ≤ −100 dB | ≤ −110 dB |
| Dark current | 260 e⁻/s at 80 °C | ≤50 e⁻/s at 60 °C | ≤20 e⁻/s at 60 °C |
| Colour-accurate range, neutral / saturated | n.p. | ≥80 / ≥70 dB | — |
| Power, 90 fps, 12-bit | n.p. | ≤400 mW | ≤350 mW |
| Chief ray | 15° | 0° | 0° |
| Frame rate | 120 fps, HDR rate n.p. | 90 fps, every mode | 90 fps, every mode |
Dark current is quoted at the temperature in the row. Parasitic light in the storage node is the pixel to watch: published charge-domain overflow pixels often sit near −80 to −89 dB, and a short exposure in a high-contrast scene wants about −100 dB. Shield variants are on the test chip. Power is a gate at engineering samples, because reading both paths costs watts. Shipping parts avoid that cost by dropping frames.
Low light
At a luminance SNR of 10, photons do the work: pitch squared, quantum efficiency, f-number, exposure. Read noise barely moves that point. It moves the “something is visible” floor by about 3×. A separate high-gain conversion is required — a single 12-bit span across a deep full well is too coarse at one lux.
Range
Brochure dynamic range is full well over the noise floor. The number that matters is the span that stays linear, at full frame rate, in the right colour. A very large overflow capacitor can print well above 110 dB and still break colour at the join. Targets stay at 80 dB and 90 dB.
No published 2.3 MP colour global shutter has shown 90 dB, single exposure, at 90 fps. Closest anchors: an onsemi 4.2 µm overflow global shutter near 90 dB at 30 fps, a Brillnics 4.0 µm stacked overflow pixel near 90 dB, a Canon charge-domain pixel at 1.8 e⁻ read noise.
Frame rate
Each row needs four samples: reset and signal, for the high-gain path and the overflow path. One converter per column runs out of time, which is why shipping dual-gain parts halve the frame rate. GS2.3 uses two converters per column and merges on the chip.
Overflow charge from the frame being read has to clear before the next exposure reuses the node. The practical HDR exposure at 90 fps is about 7.5–9 ms rather than the full 11. That costs under 2 dB of low-light SNR.
Two raw 12-bit paths are about 5 Gb/s. Merged and companded to 16 bits, about 3.3. Both fit 4-lane MIPI D-PHY once the merge is on the die.
Colour
Red, green and blue do not switch paths at the same brightness. Green usually arrives first. A white highlight with green clipped is magenta, and a training set will learn the magenta.
- Both paths are read for every pixel, every frame. The choice is digital.
- Path gain is calibrated per pixel, at test and in the field, from −20 °C to 85 °C. Gate: ≤1% error, ≤0.5% on green.
- A colour quad switches together, with a soft blend.
- The flag travels with the pixel: path, clip, calibration version.
- Overflow points can follow the measured white balance each frame. A static daylight setup drifts under tungsten.
Colour-accurate dynamic range is the span where luminance SNR stays at least 10 and no channel is clipped. Pass line, once measured: mean ΔE00 ≤2, 95th percentile ≤4, above 20 dB SNR. No step above 1 at a transition. Clipped highlights stay neutral.
Motion
One pixel of blur at 1 m/s and 1 m, on a 90° field, needs about 1.0 ms. At 5 m/s, about 0.2 ms. At those exposures a 4.5 µm pixel still wants on the order of 150 lux, and several hundred lux, for a clean colour image. Pitch, quantum efficiency and a faster lens do not close a hundred-fold gap.
SyncLink’s strobe is locked to the global exposure. A white flash at 1 m, 1 ms, clean colour: about 4.5 W average. Practical out to about 2 m. Not at 5 m. A 90 Hz visible strobe is visible to people; eye safety and flicker sit in the lamp design. Near-infrared adds no measured colour.
SyncLink
| Behaviour | |
|---|---|
| Wire | One shared line on a board, or a daisy chain. |
| Master | Elected. Another takes over if it drops. |
| Time | 64-bit counter, corrected every frame. Timestamp on every frame. |
| Cables | Delay measured and removed. |
| Exposure | Align start, centre, or end. 90 / 45 / 30 fps can lock together. |
| Rigs | Exposure and gain coordinated across a pair or a ring. |
| Absolute time | GNSS one-pulse-per-second. |
| Light | Strobe aligned to the global exposure. |
| Accuracy | ≤100 ns on a board. ≤250 ns over 5 m of cable. ≤1 µs to GNSS. |
Optics and die
0° chief ray means the bundle centre hits every pixel straight on. The cone is still ±21° at f/1.4. The lens has to be image-side telecentric within ±3°.
| 4.0 µm | 4.5 µm | |
|---|---|---|
| Active area | 7.68 × 4.80 mm | 8.64 × 5.40 mm |
| Diagonal | 9.06 mm | 10.19 mm |
| Die, single wafer | ~9.0 × 7.6 mm | ~10.0 × 8.2 mm |
| Gross dies / 300 mm, stacked | ~1,175 | ~953 |
Against the sensors in use
| SC233HGS | AR0234 | AR0235 | Mira220 | GS2.3 | |
|---|---|---|---|---|---|
| Pixel | 3.0 µm BSI | 3.0 µm | 2.8 µm | 2.79 µm stacked | 4.0 µm BSI |
| Array | 1920 × 1200 | 1920 × 1200 | 1920 × 1200 | 1600 × 1400 | 1920 × 1200 |
| Read noise | 3.24 e⁻ | n.p. | n.p. | 8.5 e⁻ | ≤2.0 e⁻ |
| Full well | ~5,900 e⁻ | n.p. | n.p. | 10,700 e⁻ | ≥25,000 e⁻ |
| Linear range | 62.65 dB | 70 dB | 65.3 dB | 62 dB | ≥80 dB |
| HDR | Knee | — | — | — | Linear, one exposure |
| Green QE | n.p. | n.p. | n.p. | n.p. (mono 94%) | ≥65% |
| Chief ray | 15° | 0° or 28° | — | 0–30° | 0° |
| Full-res fps | 120 | 120 | 120 | 90 | 90, every mode |
| Sync | n.p. | External | On-chip trigger | External | SyncLink |
| Sensor | Pitch | Read noise | Full well | Linear range | HDR at full rate |
|---|---|---|---|---|---|
| SC233HGS | 3.0 µm | 3.24 e⁻ | ~5.9 ke⁻ | 62.65 dB | Knee, compressive |
| SC136HGS | 4.0 µm | n.p. | 24.8 ke⁻ | 62.64 dB | Dual gain, half rate |
| GMAX4002 | 4.0 µm | 2.6 e⁻ | 11.6 ke⁻ | 68 dB | None |
| IMX392 | 3.45 µm | 2.37 e⁻ | 10.8 ke⁻ | 73.2 dB | None. G QE 59.9% |
| IMX421 | 4.5 µm | 2.68 e⁻ | 11.1 / 25.3 ke⁻ | ~72 dB | Gains are alternatives. G QE 68% |
| IMX900 | 2.25 µm | 5.5 e⁻ | 9.8 ke⁻ | 65 dB | Spatial, mono. G QE 73.7% |
| OG02C10 / OG05C10 | 3.45 µm | <2 e⁻ | ~20 ke⁻ | 64 dB | 77 dB at half rate or less |
| AR0234CS | 3.0 µm | n.p. | n.p. | 70 dB | None |
| GS2.3 | 4.0 µm | ≤2.0 e⁻ | ≥25 / ≥80 ke⁻ | ≥80 / ≥90 dB | One exposure, 90 fps |
The rows above are SmartSens SC233HGS and SC136HGS, onsemi AR0234, AR0234CS and AR0235, ams Mira220, OmniVision OG02C10 and OG05C10, Sony IMX392, IMX421 and IMX900, and Gpixel GMAX4002. Other sensors: vendor data and EMVA 1288 camera reports. Sony green QE is published and shows the Tier 1 colour numbers sit in range of real colour global shutters.
After GS2.3
| Part | When | |
|---|---|---|
| GS2.3 | Samples month 24, production month 30 | This sensor. |
| GS1.3, GS2.3-RGBIR | +12 months | Lower cost, and colour plus infrared. |
| GS2.3 HDR+ | +18 months | Tier 2 range, 4.5 µm option. |
| GS5.0 | +24 months | Higher resolution for mapping and capture. |
| A-series | Later | Automotive grade. |
Month 12 silicon: read noise, full well, join SNR, parasitic light at 530 nm and f/1.4, per-channel QE at 0° chief ray, colour step across each transition. Month 24: a moving scene against SC233HGS at 90 fps, four-camera sync, raw sets. Month 27: third-party EMVA 1288 on at least 30 samples.
