← All notes

Note · Display & power

Circuits for displays: pixel drivers, LEDs, switched-capacitor converters, and LDOs

Ramy Rady · October 2026 · 13 min read

A phone or watch display is a large analog circuit: millions of transistors on glass, each holding a voltage that sets a current through a light-emitting diode, refreshed dozens of times a second, and powered from rails that must stay quiet enough that the eye never sees them move. This note covers the circuits that make that work: the pixel driver, the LEDs it drives and how to model them, the driver chip that feeds the columns, and the switched-capacitor converters and LDOs that power it all.

1How a display is organized

An active-matrix display is a grid. Each row has a gate (scan) line and each column a data (source) line. One row at a time, the gate line turns on that row's switch transistors, every column writes its data voltage into the selected pixels, and the row turns off and holds the value on a storage capacitor until the next frame.

t_{line} \approx \frac{1}{f_{frame} \times N_{rows}} 2400 rows at 120 Hz \Rightarrow t_{line} \approx 3.5 µs

(1)

Everything in the column path, the driver's output buffer, the long resistive-capacitive data line, and the pixel's switch, must settle inside that line time, minus blanking and compensation phases.

LCD: the liquid crystal is a capacitor whose voltage sets how much backlight passes. One transistor and one capacitor per subpixel (1T1C); voltage-driven.
OLED and micro-LED: each subpixel emits its own light in proportion to current. The pixel circuit must turn a stored voltage into an accurate current for a whole frame.

2Thin-film transistors

The pixel transistors are thin-film transistors (TFTs) deposited on glass or polyimide. Their properties set what the pixel circuit has to correct:

Technology	Mobility (cm²/V·s)	Strength	Weakness
Amorphous silicon (a-Si)	≈ 0.5–1	Uniform, cheap, large area	Slow; threshold drifts under bias stress
Low-temperature polysilicon (LTPS)	≈ 50–100+	Fast, CMOS (n and p)	Grain boundaries make the threshold vary from TFT to TFT
Oxide (IGZO)	≈ 10–20	Very low leakage, uniform	n-type only; threshold shifts with light and bias
LTPO (LTPS + oxide)	both	LTPS drivers with oxide switches	More process steps

LTPO is what lets watches and phones drop the refresh rate to 1–10 Hz for static content: an oxide switch leaks so little that the stored voltage barely moves over a long frame.

3The 2T1C pixel

The simplest current-driving pixel has two transistors and one capacitor. The switch T1 connects the data line to the storage capacitor C_st while the row is selected. The drive transistor T2 then converts the stored gate voltage into the LED current, operating in saturation:

I_{LED} = \frac{1}{2} μ_{eff} C_{ox} \frac{W}{L} {(V_{GS} - V_{th})}^{2}

(2)

Figure 1. A 2T1C pixel (drawn with an n-type drive transistor). The storage capacitor holds T2's gate voltage for the whole frame after T1 turns off.

The square law is the problem. A threshold error $Δ V_{th}$ changes the current by

\frac{Δ I}{I} \approx - \frac{2 Δ V_{th}}{V_{GS} - V_{th}}

(3)

With 1 V of overdrive, a 50 mV threshold difference is a 10% brightness difference. The eye notices a few percent between neighbouring areas as blotchy non-uniformity (“mura”), and dark gray levels, which use small overdrive, are the worst.

Other errors add to the threshold spread: threshold drift with time (bias stress in a-Si and oxide), IR drop along the ELVDD supply mesh so pixels far from the feed see less V_GS, OLED aging that raises the diode's voltage and lowers its efficiency, and charge injection and clock feedthrough from T1 when it turns off.

4Compensating the pixel

Internal compensation: sample the threshold inside the pixel

Most LTPS phone OLEDs use a pixel with around seven transistors and one capacitor (7T1C). With a p-type drive transistor, the sequence is:

Initialize: reset the drive transistor's gate and the LED anode to known voltages.
Sample and program: connect the drive transistor as a diode while the data voltage is applied to its source. Its gate settles to $V_{data} - | V_{th} |$ , storing the threshold on C_st.
Emit: switch the source to ELVDD. Now $V_{SG} = V_{ELVDD} - V_{data} + | V_{th} |$ , and the threshold cancels:

I_{LED} = \frac{k}{2} {(V_{SG} - | V_{th} |)}^{2} = \frac{k}{2} {(V_{ELVDD} - V_{data})}^{2}

(4)

The cost is time: the diode-connected transistor must settle within the line time, and that slows down at low currents. A separate emission-control transistor also lets the panel turn rows off for part of a frame, which is how many OLEDs dim with PWM.

External compensation: measure and correct in the driver

Large OLED TVs add a sense transistor and sense line to each subpixel. The driver periodically measures each drive transistor's current (and the OLED's voltage), and corrects the data value digitally. It handles threshold and mobility variation and slow aging, at the cost of an extra line per column and a sensing ADC in the driver.

Current programming

Programming the pixel with a current instead of a voltage corrects threshold and mobility at once, but small currents charge the long data line slowly, so dark gray levels settle too late. Chaji's thesis (reference 1) develops faster schemes for exactly this problem.

5LEDs and how to model them

For circuit design, an LED (organic or inorganic) is a diode with a series resistance, a small leakage path, and a capacitance:

I = I_{S} (e^{(V - I R_{S}) / n V_{T}} - 1) + \frac{V - I R_{S}}{R_{P}}

(5)

Turn-on voltage follows the photon energy: roughly 1.8–2.0 V for red, 2.6–3.0 V for green and blue InGaN. The ideality factor $n$ is often 1.5–3 in real devices.
The capacitance (an OLED is a thin film between two electrodes, often several femtofarads per square micron) has to be charged to the turn-on voltage before light appears, which matters for short PWM pulses and low gray levels.
Efficiency splits into how many electrons make photons (internal quantum efficiency, IQE) and how many photons escape (light extraction): $EQE = IQE \times η_{extraction}$ .

The standard model for IQE is the ABC model: carriers recombine through defects (rate $A n$ ), radiatively ( $B n^{2}$ ), or by Auger processes ( $C n^{3}$ ). Only the middle term makes light:

IQE = \frac{B n^{2}}{A n + B n^{2} + C n^{3}}, A_{eff} \approx A_{0} + \frac{4 S}{d}

(6)

For a square micro-LED of side $d$ , damaged sidewalls add non-radiative recombination with surface velocity $S$ , in proportion to perimeter over area, $4 / d$ . Shrinking the LED raises $A$ , and efficiency at the low current densities a display uses falls sharply. Sidewall passivation (reference 4) recovers much of it.

6The display driver IC

A mobile display driver IC (DDIC) takes video over MIPI DSI (or eDP for laptops), stores or buffers it, and drives a thousand or more data columns. Its main blocks:

Timing controller: generates the scan and emission timing, and on many panels drives a gate driver built from TFTs on the glass itself (gate-on-array), a shift register that uses bootstrapped capacitors to swing its outputs past the supply.
Gamma: maps 8–10 bit codes to voltages on the panel's nonlinear brightness curve, usually with a tapped resistor string per color.
Source amplifiers: a rail-to-rail class-AB buffer per column that drives the data line's distributed resistance and capacitance. Its offset must be a fraction of one gray step, so drivers use offset cancellation or chopping between frames.

Settling sets the speed requirement. To settle within half an LSB of an N-bit level through a single time constant τ:

t_{settle} = τ \ln 2^{N + 1} N = 10 \Rightarrow t_{settle} \approx 7.6 τ

(7)

With a few microseconds per line and part of it used by compensation, the column's RC and the buffer's slew rate leave little margin, especially at the far end of a long data line.

7Power: rails and converters

Rail	Typical level	Feeds	Usually made by
ELVDD	about +4 to +5 V	Drive transistors and LEDs (positive side)	Inductive boost converter
ELVSS	a few volts negative	LED cathodes (common)	Inverting buck-boost
AVDD	about 5–7 V	Source amplifiers, gamma	Boost or charge pump
VGH / VGL	several volts above and below	Gate driver high and low levels	Charge pumps
Analog references	precise, quiet	Gamma reference, bias	LDOs after the converters

The panel current depends on the image: a white screen can draw several times what a dark one does. ELVDD and ELVSS must therefore hold their voltage across large, fast load steps, because the voltage across every pixel depends on them. Ripple shows up directly as flicker or banding.

8Switched-capacitor converters

A switched-capacitor (SC) converter moves charge with capacitors and switches only, no inductor, which makes it small and easy to integrate. Its ideal output is a fixed ratio of the input (2:1, 1:2, 3:2, …). Under load the output drops as if through an output resistance, which has two limits (Seeman and Sanders, reference 5):

R_{SSL} = \sum_{i} \frac{a_{c, i}^{2}}{C_{i} f_{sw}}, R_{FSL} = 2 \sum_{i} R_{i} a_{r, i}^{2}, R_{out} \approx \sqrt{R_{SSL}^{2} + R_{FSL}^{2}}

(8)

Slow-switching limit (SSL): the capacitors fully charge and discharge each cycle; loss comes from charge sharing and falls as $1 / (C f_{sw})$ . The $a_{c}$ are each capacitor's charge per unit of output charge.
Fast-switching limit (FSL): the capacitors barely change voltage; loss is set by switch resistance, independent of frequency.
For a 2:1 series-parallel converter with one flying capacitor and four switches, $R_{SSL} = 1 / (4 C f_{sw})$ and $R_{FSL} = 2 R_{on}$ . Switching faster than the corner where they meet only adds gate-drive loss.

9LDO design

A low-dropout regulator (LDO) is an error amplifier, a pass transistor (usually PMOS so the dropout can be a few hundred millivolts), and a feedback divider. Display drivers use LDOs after the switching converters to give the gamma references and analog blocks a supply free of switching ripple.

Figure 2. An LDO regulator. The loop holds the divided output equal to the reference: V_OUT = V_REF(1 + R₁/R₂).

The specifications and the equations behind them:

Specification	What sets it
Dropout	Pass device on-resistance at full load: $V_{DO} = I_{max} R_{on}$
Load regulation	$Δ V_{out} / Δ I_{L} \approx R_{out,open} / (1 + T)$ , with loop gain $T$
Load transient	$Δ V \approx Δ I_{L} Δ t / C_{out}$ , where $Δ t$ ≈ loop response time (about $1 / {BW}_{cl}$ ) plus gate slewing time
PSRR	At low frequency, roughly the loop gain; above the loop bandwidth, the output capacitor and the pass device's parasitics
Quiescent current	Error amplifier and divider bias; a tradeoff against speed

Stability

The loop has at least two poles: the output pole $1 / (2 π R_{L} C_{out})$ , which moves by orders of magnitude with load current, and the pole at the large gate of the pass device. Two families of solutions:

External capacitor (µF): the output pole is dominant and the loop's unity-gain frequency is $\approx g_{m,loop} / (2 π C_{out})$ . The gate pole must sit above it; historically the capacitor's ESR zero helped.
Capacitor-less (on-chip, pF to nF): the output pole is fast, so the gate node is made dominant with Miller compensation and techniques such as damping-factor control (reference 6). These LDOs settle in nanoseconds but rely on a fast loop for transients.

Phase margin follows from where the second pole sits: $PM \approx 90^{\circ} - \arctan (f_{UGB} / f_{p_{2}})$ .

Questions & answers

Why are OLED pixels current-driven when LCD pixels are voltage-driven?

A liquid crystal responds to the voltage across it and draws almost no current, so storing a voltage is enough. An OLED's brightness is proportional to the current through it, and its voltage for a given current varies from pixel to pixel and with age, so the pixel must turn a stored voltage into a controlled current.

Why are dark gray levels the hardest to make uniform?

From eq. (3), the brightness error scales with ΔV_th divided by the overdrive. Dark levels use small overdrive, so the same threshold error is a larger fraction. In the panel simulation above, lower the gray level and watch the spread grow.

When would I choose a switched-capacitor converter over an inductive one?

When the conversion ratio is fixed or nearly fixed, the current is modest, and size or integration matters, such as gate-driver rails or a 2:1 step-down on chip. For wide ratios, high currents, or tight regulation over a wide input range, an inductive converter is more efficient.

Why put an LDO after a switching converter instead of just using the converter?

Switching converters leave ripple at their switching frequency and its harmonics. An LDO with good PSRR at those frequencies removes it, giving references and amplifiers a quiet supply, at the cost of the dropout voltage times the load current in efficiency.

Where can I go deeper on pixel circuits?

Start with the review by Nathan, Chaji, and Ashtiani (reference 2), then Chaji's thesis (reference 1) for the circuits and their analysis. For micro-LEDs, follow the SID Digest papers on PWM pixel circuits (reference 3).

1How a display is organized✎Edit

2Thin-film transistors✎Edit

3The 2T1C pixel✎Edit

4Compensating the pixel✎Edit

Internal compensation: sample the threshold inside the pixel

External compensation: measure and correct in the driver

Current programming

5LEDs and how to model them✎Edit

6The display driver IC✎Edit

7Power: rails and converters✎Edit

8Switched-capacitor converters✎Edit

9LDO design✎Edit