Global shutter CMOS sensor comparison: speed vs dynamic range

Global shutter made CMOS the default for moving-part inspection, but it did not make the sensor choice simple. Inside global shutter CMOS, two families spend the same die area on different goals: one buys readout lanes and frame rate, the other buys larger pixels, deeper full wells and quieter readout. This comparison gives the numbers that pick a side — dynamic range from full well and read noise, frame rate from readout and windowing — then walks one camera through focal length, working distance, exposure and f-number, so you can order with a decision instead of a hunch. The CCD versus CMOS differences explain why CMOS won; the question here is which CMOS.

The decision first

Pick the family your binding constraint punishes least.

• Scene needs more than about 70 dB of usable range — glare beside near-black, weld pools, backlit glass — take the dynamic-range-first family: larger pixels, full wells in the 10,000-30,000 electron class, read noise of 2-3 electrons, 70-84 dB computed.
• Line needs frame rate at resolution — more than roughly 60 fps at 3 MP and above, or several cameras sharing one cable — take the speed-first family: 3-4 µm pixels, 5,000-10,000 electron wells, 58-70 dB.
• Neither constraint bites. The default is a 2.3 MP-class global shutter at 40 fps, windowed to SXGA, XGA or VGA when you need rate. Spend on a specialized part only when one of the thresholds above fires.

Both thresholds reduce to arithmetic you can repeat on any data sheet: scene contrast in dB against 20 × log10(full well ÷ read noise), and line rate against full-frame readout. The next two sections run each sum.

What readout speed buys

A global shutter pixel parks its charge in an on-chip hold; the chip then converts the frame and pushes it off-sensor. The full-frame period is close to pixel count divided by however many convertor lanes the designer laid out, plus reset overhead. Lanes and fast convertors are silicon area, and so are big pixels. That is the entire trade: same die, different spending.

Take the general-purpose class: a 2.3 MP global shutter reads 40 fps at full frame. Rate scales with the rows you actually read, so windowing the same silicon to SXGA (1024 rows) gives about 40 × 1200 ÷ 1024 = 47 fps, XGA (768 rows) about 40 × 1200 ÷ 768 = 63 fps, and VGA (480 rows) about 40 × 1200 ÷ 480 = 100 fps. Width cuts help CMOS readouts further when the sensor is row-limited — camera ROI and frame rate covers that — and reading several regions in one frame is the multi-ROI readout case. Choosing among standard window sizes changes the shutter behavior too; selecting WVGA or SXGA imaging modes walks one camera's version of that choice.

Board-level parts lead speed budgets because small formats read out fastest. The LS-BCM-0003 is a 3.1 MP USB3 board camera on a 1/1.8" global shutter sensor at 120 fps with a C-mount. Binning trades resolution for rate or sensitivity when you are close; 4x4 versus 2x2 binning and frame rate quantifies it, and sensor selection for high-speed inspection starts at the same readout numbers.

What dynamic range is

Dynamic range is the ratio between the most charge a pixel can hold — full well, in electrons — and the noise floor in the dark — read noise, in electrons. On a decibel scale:

SNR_max is the shot-noise ceiling at saturation: 10,000 electrons gives 100:1, which is 40 dB; 15,000 gives 122:1, which is 41.8 dB. Bigger wells buy quieter bright signals, which is why photometry and measurement work favors the range-first family.

The same well moves a long way with noise: 10,000 electrons over 2.5 electrons of read noise is 72.0 dB; over 6 electrons it is 64.4 dB. Nothing about the pixel changed — the readout did.

Sensor data sheets and EMVA 1288 sheets print full well and read noise per gain setting. To find what your scene needs, take one raw frame at fixed gain and read the gray-value ratio between the brightest region that must keep detail and the darkest — or do it properly with a gray-step target measurement. A polished bracket on black foam measures about 2000:1: 20 × log10(2000) = 66 dB, plus 6 dB of margin for one bit, so 72 dB needed. A speed-first part computing 6,000 ÷ 3.5 = 1714, which is 64.7 dB, clips the highlight end. A range-first part computing 15,000 ÷ 2.4 = 6250, which is 75.9 dB, holds it. One line of arithmetic, family decided.

The two families side by side

Recompute the dB row from the full well and read noise of the exact part before you commit; the brackets are class-typical for current global shutter CMOS, and individual parts sit anywhere inside them.

Three catalog parts frame the spread. The MV-CU020-90GC is a 2.5 MP color GigE camera on a 1/1.7" global shutter sensor — the general-inspection default. The LS-BCM-0003 above is the embeddable speed case. The MV-CH120-11UC is a 12 MP USB3 camera on a 1.1" global shutter sensor in a high-speed version — detail and range together.

All 3 share: Lens mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Worked example: a 3.1 MP global shutter over a 64 mm tray at 250 mm

Inputs, each with its source:

• Sensor: 2064 × 1544 pixels at 3.45 µm, from the data sheet of the sensor family behind the 3.1 MP board camera above. Cross-check: 2064 × 3.45 µm = 7.12 mm of sensor width, matching the 1/1.8" format class the catalog lists.
• Field of view needed: a 64 × 48 mm tray.
• Working distance: 250 mm, set by the machine guarding.
• Tray speed: 120 mm/s, continuous past the camera.
• Smallest feature to catch: 0.25 mm.
• Light: strobe allowed.

The scene numbers are example inputs; swap yours in and the arithmetic holds.

Rules in words: magnification is sensor width over field of view; focal length follows from working distance and magnification; the exposure ceiling comes from motion blur; the f-number from depth of field.

The sums:

• m = 7.12 mm ÷ 64 mm = 0.111.
• f = 250 mm × 0.111 ÷ 1.111 = 25.0 mm, so a 25 mm C-mount lens at the 250 mm station.
• Scale: 64 mm ÷ 2064 px = 31 µm per pixel. The 0.25 mm feature spans 250 ÷ 31 = 8 px, comfortably over the 3-5 px minimum for dependable detection.
• Exposure: half a pixel is 15.5 µm, so t_max = 15.5 µm ÷ 120 mm/s = 130 µs. The frame period at 120 fps is 8.3 ms; the exposure fits 64 times over, so readout is not the binding constraint — light is.
• f-number: the depth-of-field estimate 2 × N × c × (1 + m) ÷ m², with c the 3.45 µm pixel, gives about 2.5 mm total at f/4 and 5.0 mm at f/8. Tray height varies ±1 mm, so f/4 holds; f/8 adds depth but costs two stops — one quarter of the light — which you do not have at 130 µs.

Feed your own tray numbers to the focal-length calculator:

A 130 µs ceiling puts you in strobe territory. Pulse a high-current LED so the light pulse, not the integration window, freezes the tray; the global shutter holds every row open for the same pulse. Microsecond laser and shutter control over high-speed I/O shows how close the pulse and the exposure must sit.

Decision: the tray scene is evenly lit with contrast around 100:1 — 20 × log10(100) = 40 dB, plus margin, 46 dB needed — inside both families. The binding constraint is the 130 µs exposure, so the fast small-pixel readout wins and the budget goes to light, not well depth. Put chrome on black foam on the same tray and the needed range jumps past 70 dB: same geometry, other family.

Which family for which job

Fast lines, more than 60 fps at 3 MP and up. Speed-first, with interface to match. For large trays at rate, the MV-CH250-25TC-F-NF reads a 25 MP APS-H sensor over 10GigE through an F-mount and ships with a fan — big readouts dissipate real heat.

Glare-to-black scenes past 70 dB. Range-first. The MV-CH120-11UC is a 12 MP USB3 camera on a 1.1" global shutter sensor in a high-speed version — the format that keeps detail and headroom together.

General bench inspection. No strobe, no glare, moderate rate: the general-purpose class from the table above, and the MV-CU020-90GC card below.

Embedded builds. Board-level cameras answer to the same two thresholds; the 120 fps part in the speed section is the reference for a small, fast readout.

On-device inspection. Smart cameras and vision sensors come with the sensor fixed by the platform, so run both thresholds against the listed sensor before committing the platform. The same 70 dB and 60 fps lines decide it.

Checks before you order

1. Lens against pixel. At 3.45 µm pitch the Nyquist limit is 1 ÷ (2 × 3.45 µm) ≈ 145 lp/mm; the lens must hold contrast near there and its image circle must cover the format. Why smaller pixels give lower contrast and what formats like 1/1.8" mean cover both failure modes.
2. Cable against readout. 2064 × 1544 px × 120 fps × 1 byte = 382 MB/s, about 3.1 Gbit/s before overhead — USB3 or 10GigE territory, not a 1 Gbit/s link. 5GBASE-T bandwidth and distance is the middle option.
3. Bits end to end. 8-bit processing throws away the top 24 dB of a 72 dB sensor. Keep 10- or 12-bit output into the algorithm; when 12-bit data lands in 16-bit containers, why the values top out near 65520 saves an afternoon of debugging.
4. Auto functions off. Auto exposure and gain quietly cap the frame rate; turning them off restores the rated fps.
5. Exposure before frame rate. On moving parts, exposure time versus shutter speed sets what you can freeze; the frame rate only sets how often.
6. Heat path. A fan on a 25 MP camera is a hint: plan airflow for the enclosure, not just the sensor.

When the usual answer is wrong

• Nothing moves. With no motion to freeze, the hold node buys you nothing for the sensitive area it consumes; a rolling shutter sibling of the same family converts more photons and reads quieter at the same price. For static scenes, take it; global versus rolling shutter for moving targets draws the line.
• You can flash. A rolling sensor with a global-reset flash freezes motion for less money — rolling shutter global reset with flash is the trick.
• Flat lighting, 8-bit pipeline. The extra range never reaches the algorithm; buy rate or price instead.
• Speed-first behind a slow link. The wire sets the frame rate. Window the image instead of paying for readout you cannot export.
• Range-first on a fast line. You spent as frame rate the headroom you actually needed.

Checklist before ordering

• Scene range: 20 × log10(contrast ratio) + 6 dB; above 70 dB, range-first.
• Line rate: needed fps at needed resolution against the full-frame rate; try row windowing first.
• Lens: resolves near 1 ÷ (2 × pixel pitch); image circle covers the format.
• Link: bytes per second × 1.25 for overhead, against the cable's rate.
• Exposure: t_max from half-pixel blur, with a light and f-number that deliver it.
• Pipeline: bit depth kept from sensor to algorithm; auto functions off.
• Mechanics: airflow for high-duty readout; guarding or interlock around the strobe.

When the thresholds tie, take the lower read noise at the gain you will run. Then start at the area scan cameras finder filtered to Shutter: Global, or let the camera configurator walk the whole chain. Two parts still tied after the sums? Send us the scene and we will run the numbers with you.