블로그 약 Is a 4.3-inch High Brightness 1000nits TFT LCD Still a Microcontroller Project?

The SFTC240H-style modules that most engineers meet first carry a controller with its own frame memory. You write pixels when you feel like it, the panel holds them, and the display keeps scanning without you. The SFTO430JS-7140B-N from Saef Technology Limited is a 4.3-inch High Brightness 1000nits TFT LCD and works differently: 480 x 272 pixels on an IPS cell, 1000 cd/m² typical, 800:1 contrast, 6.6% transmittance, a 24-bit RGB parallel interface driven by an NV3047F on a 40-pin FPC, 105.5 x 67.2 x 3.0 mm, and a backlight that draws 1.024 W. The controller has no frame memory. It is a receiver, not a storage device, and that single fact reorganises everything around it.

Three numbers carry the argument: 255 KB, 27 MB/s and 58 microseconds. Each one is a threshold your design crosses the moment you choose this class of panel, and none of them appears in a feature table.

255 KB: the frame buffer decides the microcontroller

Start with the arithmetic, because it is short. 480 x 272 is 130,560 pixels. Stored as RGB565 that is 261,120 bytes, or 255 KiB. As RGB888, which is what this module's 24-bit bus is built for, it is 391,680 bytes, or 383 KiB. Double-buffer either of them for tear-free animation and you need 510 KiB or 766 KiB of contiguous, always-addressable memory.

Now compare that with what sits inside the microcontrollers engineers actually specify:

The STM32F407 row is not a memory verdict. That part has no RGB display controller at all: its only display peripheral is an 8080/6800 bus, so it cannot source this interface at any clock rate, with any amount of memory hung off it. Keeping the F407 means adding an external RGB controller, which is another way of buying back the frame memory the panel does not contain.

Two qualifiers make the table stricter than it looks. The first is contiguity: a display controller such as the STM32 LTDC or NXP LCDIF fetches from one linear region, so 320 KB spread across five SRAM banks is not the same as 320 KB in one block. The second is the double-buffer row. On the STM32H743, whose 512 KB AXI SRAM is the most generous contiguous block in this class, two RGB565 buffers consume 510 of those 512 KB and leave nothing for the application.

There is a third trap that catches people who have done the arithmetic correctly. On an STM32H7 the fastest memory is the tightly-coupled RAM, and neither the LTDC nor the DMA engine can reach it. The buffer therefore has to live in the 512 KB AXI SRAM at 0x24000000, which is the only large block the display controller can address, and which then becomes unavailable for anything else. The memory that is fast enough is not the memory that is reachable, and the memory that is reachable is the memory you wanted for the application.

So in practice the frame buffer goes to external memory: an SDRAM on a parallel bus, or octal PSRAM on an ESP32-S3. Both work. Both also introduce a second clock domain, a second set of timing constraints, and a memory whose bandwidth you now share with everything else in the product. That is the first reclassification: the choice of display has effectively chosen your memory architecture, and with it a large part of the schematic.

27 MB/s: a bus that never stops

The second number is what the interface does with that memory. At a 9 MHz DCLK and 24 bits per pixel, the bus clocks 27 MB/s continuously, of which about 23.5 MB/s is useful pixel data and the rest is blanking. For comparison, the 8-bit 8080 bus on a small controller panel peaks around 15 MB/s and the fastest practical four-wire SPI around 8 MB/s — but both of those are burst interfaces that idle when you are not writing. An RGB bus has no idle state. It runs from power-up to power-down, whether or not anything on screen has changed.

Twenty-seven megabytes per second is not fast by modern standards, and the signal integrity at 9 MHz, a 111 ns period, is undemanding. What makes this interface hard is the line count and the return path. Twenty-eight signals, 24 of them data, leave the module on a 40-pin 0.5 mm FPC that provides three ground pins: one ground for every nine signals. That ratio is the real EMC problem, and it is a layout problem rather than a firmware one: each data line needs a 22-33 Ω series resistor at the driver, the DCLK line needs ground on both sides for its whole length, and the FPC tail should stay short. How short is answerable: at a 9 MHz clock the knee is near 70 MHz, a fifteenth of that wavelength is 285 mm, so a 150 mm tail needs no controlled impedance, and the series resistors are there for EMI rather than for signal integrity. When a connector offers three grounds for twenty-eight signals, with up to twenty-four consecutive pins carrying data and no return between them, the ground return has to be recreated on the host PCB as close to the connector as the footprint allows, and the two pins that carry the highest dI/dt, DCLK and the LED string return, deserve their own via field rather than sharing one.

There is an upside to the low pixel clock that is easy to miss. Because 9 MHz is slow, you have enormous timing margin: the datasheet's 12 ns data setup and hold requirements are met with an 11% budget against a 111 ns cycle, so a modest microcontroller's GPIO or a small FPGA can source this bus without dedicated hardware. The difficulty is not speed. It is that the speed never reaches zero.

That idle-free bus also explains why the power sequence on this module class is specified in hundreds of milliseconds rather than in microseconds. The signal lines have to be stable for at least 160 ms before the backlight is enabled, and the backlight has to be off for at least 160 ms before the signals are removed. The reason the module specification gives is to prevent latch-up and DC operation of the cell, which in plain terms means driving an unpowered panel back-powers its own rails through the ESD diodes and leaves a standing voltage across the liquid crystal. Neither number matters on a controller panel, where the display holds its last frame and the backlight can be switched whenever convenient. Here the RGB stream is the only thing that keeps the panel defined, so a warm reset with VDD still standing leaves it scanning whatever arrives until you gate DISP and restart the timing. Sequence that enable in hardware, tied to the backlight rail, rather than in the reset handler.

58 microseconds: the deadline moves from frame to line

This is the number that changes how the firmware is written, and it comes straight out of the AC characteristics table rather than from a marketing summary.

The datasheet specifies an HSYNC period of 55 to 65 µs, a clock duty cycle of 40 to 60%, and 12 ns setup and hold on data, DE, HSYNC and VSYNC. It does not state a maximum DCLK frequency, because the HSYNC period is the constraint, and that constraint is more useful than a headline megahertz figure. With the conventional 480x272 total of 525 clocks per line, 55 to 65 µs brackets DCLK between 8.1 and 9.5 MHz, and the familiar 9 MHz / 525 x 286 combination sits inside it at 59.9 Hz with a line period of 58.3 µs. The window is wide enough to absorb a PLL that cannot land on 9.000 MHz exactly, which most cannot. It is not wide enough to absorb a designer who trims the porch values. Those 525 clocks are a floor rather than a convention: 480 active plus 41 HSYNC plus 2 and 2 of front and back porch, and the source drivers need at least 44 clocks of blanking to settle. Cut into that to buy refresh rate and the panel does not refuse to run, it simply begins each line before its source outputs have arrived, which shows as a vertical band of wrong colour along one edge.

Fifty-eight microseconds is the time between two consecutive lines. On a controller-based panel your deadline was one frame, 16.7 ms, and the panel's own memory absorbed any lateness. Here the deadline is 58 µs and there is nothing on the panel to absorb lateness. Miss it and the line does not wait.

What fits inside 58 µs is not much. A FreeRTOS context switch is 1-10 µs, so one is survivable. An I²C transaction to a sensor is 100 µs to 1 ms, so one is not. An SD card sector write is milliseconds, and a flash page erase is tens of milliseconds, so neither is remotely survivable. Any of them arriving during an active line will produce a visible horizontal streak if the display's pixel source is the CPU.

With a full frame buffer in external memory the display controller keeps streaming autonomously, which is why the framebuffer in the previous section is not an optional convenience. But external memory brings its own deadlines. SDRAM refresh cycles steal bandwidth on a schedule you do not control, other bus masters compete for the same pins, and an underrun shows up on screen as a horizontal line of garbage at the exact moment the system was busy. Most display controllers flag it: the STM32 LTDC raises an underrun interrupt, and on an ESP32-S3 the RGB peripheral can be configured to raise one as well. If you are seeing those interrupts in the field rather than in the lab, you are looking at a bus architecture problem, and the fix is usually memory bandwidth rather than display driver code.

It helps to know how this actually presents, because the symptom does not look like a bandwidth number. A published case on an STM32H7 driving exactly this panel class, 480 x 272 at 24 bits and an 8 MHz pixel clock with two framebuffers in external FMC SDRAM, produced LTDC FIFO underrun interrupts against a static image. The display was not at fault. The Cortex-M7 was speculatively reading unused FMC address space onto the bus, and the fix was an MPU region that marked that space device memory. Two layout rules follow from the way the controller fetches. It reads in 64-byte bursts, and an SDRAM burst cannot cross a 1 KiB boundary, so the stride gets padded: 480 pixels at three bytes is 1,440 bytes, which has to become 1,472. At 27 MB/s one of those bursts is issued every 2.4 µs, about 6,100 times a frame, so a stride that is wrong once is wrong on every line rather than occasionally.

The practical consequences are specific. Give the display DMA the highest bus priority and prove it with a logic analyser on the memory bus, not by reading the reference manual. Put the framebuffer in the fastest memory available and accept that its location constrains where the rest of the application can live. And treat every long-running peripheral operation, whether storage, network or motor control, as a display event until measurements prove otherwise.

1.024 watts: eight LEDs in series and a 25.6 V rail

The backlight of this 4.3-inch High Brightness 1000nits TFT LCD is specified at 25.6 V and 40 mA, which is 1.024 W and, almost certainly, eight white LEDs in series at about 3.2 V each. That topology has a consequence worth stating plainly: the string is a single electrical path, so one LED failing open extinguishes the entire backlight. Whatever driver you choose needs open-LED protection, and on a product with a warranty period it needs it as a requirement rather than a feature.

Feeding 25.6 V from an industrial rail is the second decision. From 12 V a boost converter runs at 53% duty cycle, which is comfortable. From 24 V the same converter runs at 6% duty, and worse, it cannot regulate at all at the top of the rail: the string voltage falls to about 24.5 V at the +70 °C top of the module's operating range while a 24 V bus is allowed to reach 26.4 V, so for part of the operating envelope the input sits above the output and the boost's inductor-to-diode path conducts directly. On a 24 V system the honest options are a buck-boost, a 30 V intermediate rail feeding a linear current sink, or, the option people forget, a different LED topology. Eight LEDs in series is the default configuration, not a law of nature. A two-parallel, four-series string at 6.4 V and 80 mA suits a 12 V rail far better, and a four-parallel, two-series string at 3.2 V suits a 5 V rail. Topology is a build option, and it is worth specifying on the purchase order rather than working around in hardware.

The thermal picture is less alarming than the electrical one. 1.024 W spread over 5,119 mm² of active area is 200 W/m², and the equivalent figure for a 2.4-inch 1000-nit module at 0.496 W is 193 W/m². So the two are within 5% of each other: there is no per-area advantage to the larger panel, and no penalty either. One watt is simply a small number. In a sealed 200 x 150 x 80 mm polycarbonate enclosure it raises the internal air by about 0.8 K, and even accounting for the internal air-to-wall convection resistance the honest total is 2 to 3 K. The sun is the larger term: the 95.04 x 53.86 mm window intercepts 4.9 W of solar irradiance at an absorptivity of 0.95, roughly five times the backlight, and it does so whether the display is switched on.

Two quality metrics on this datasheet deserve attention at this size. Uniformity is 75% minimum across the active area, which on a 95 mm edge-lit light guide means the corner nearest the LED entry can sit 25% below the brightest point. That is within specification and still visible on a flat white background, so if the product will show gradients or a full-white screen, ask how uniformity was measured and at how many points. And the LED lifetime is given as 50,000 hours to 50% of initial brightness — a better figure than the 20,000 hours typical of smaller panels, because there is more emitter area for the same luminance, so each die runs at a lower current density. It is still an L50 number. If 50,000 hours is quoted at 25 °C, which is the convention, then at a 70 °C junction the same curve gives roughly 6,500 hours, so the headline figure describes a laboratory and not a cabinet. Ask for the measurement temperature before you put the number in a datasheet of your own, and treat adaptive dimming as the cheapest way to buy years.

No touch fitted: the decision that is now yours

This module ships without a touch panel. Five pins, 35 and 37 to 40, are unconnected, and there is no resistive or capacitive layer in the stack. On a 2.4-inch display that is often the right answer, because a small screen usually presents three or four numbers to one operator. On a 4.3-inch High Brightness 1000nits TFT LCD it almost never is: at this size the product is expected to accept input, and the question is not whether to add touch but which kind, and whether the mechanical envelope was designed for it.

The thickness budget is the part teams underestimate. The module is 3.0 mm. A capacitive stack with a 1.1 mm chemically strengthened cover glass over a 0.125 mm ITO film, air-gapped half a millimetre above the LCD, brings the assembly to 4.73 mm; the same stack optically bonded with a 0.175 mm adhesive comes to 4.40 mm. Ask for a 2 mm cover glass to meet an impact rating and you are at 5.88 mm, and 3 mm puts you at 6.88 mm. A housing drawn to 3.0 mm is between 47% and 129% too shallow, and the seal plane has moved forward with it.

Then the harder constraint, which is where retrofits actually fail. The bezel opening is 98.36 x 56.83 mm against a 95.04 x 53.86 mm active area, so the border available for a gasket is 1.66 mm at the sides and 1.49 mm at the top and bottom, and the opening itself carries a ±0.2 mm tolerance. A closed-cell bead needs 1.0 to 1.2 mm, bezel flatness and position tolerance take another 0.10 to 0.15 mm, and the cover glass edge tolerance takes 0.2 to 0.3 mm. Add those up and the top and bottom border is negative before you have allowed a single tenth of a millimetre for bead compression. There is no perimeter seal available inside this opening. The options are to enlarge the bezel window so the gasket lands on the cover glass print rather than on the module, to bond only the perimeter and accept a lower IP rating, or to redesign the front as a bonded display-and-touch cassette with the seal outside the optical stack. Whichever you choose, it is a housing change, and it belongs in the concept phase rather than in the ECN after tooling.

What bonding buys, beyond depth, is worth quantifying because it is usually the deciding argument. An air-gapped stack reflects about 8% at the three glass-air interfaces; a bonded stack with an index-matched adhesive brings internal reflection under 1%. Reflected luminance is E·ρ/π, so for this 1000 cd/m², 800:1 panel:

Bonding alone is worth roughly six times the apparent contrast at 10,000 lux. It also removes the half-millimetre parallax that makes a finger look offset from the cursor at the edge of the screen, and it eliminates the internal air cavity where condensation forms. What it costs is repairability: a bonded stack is scrap, not a field-replaceable part, and an OCA or LOCA layer that delaminates across the −30 to +80 °C storage range this module is rated for is a warranty event with no fix.

Electrically, a 95 x 54 mm sensor is unremarkable. Roughly eighteen drive and twelve sense channels at a 5 mm pitch, one controller — an FT6336 or GT911 over I²C with interrupt and reset, five touch points, a report rate around 100 Hz. The GT911's design rules are worth quoting to a sensor vendor because they are the ones that get missed at this size: drive-trace impedance under 3 kΩ, sense-trace under 10 kΩ, node capacitance under 4 pF, sensing RC under 6 µs. A 4.3-inch sensor with runs 40% longer than a 2.4-inch one still meets those numbers with ordinary ITO, but the margin is real and it is the vendor's job to prove rather than assert it.

Industry practice on cover glass at this size converges on 1.1 to 3 mm, and the impact ratings explain the spread: IK07 is 2 joules, IK08 is 5, IK09 is 10 and IK10 is 20, and roughly 3 mm of thermally tempered or chemically strengthened glass reaches IK08 while 6 mm is needed for IK10. That creates a genuine conflict, because 6 mm is at or beyond the point where a capacitive sensor stops seeing a gloved finger. For an outdoor unit that needs both, 3 mm chemically strengthened glass over a bonded sensor is the workable compromise, and it is a decision to take with the mechanical engineer rather than after them.

Glove performance has hard published limits. An automotive-grade TrueTouch controller is specified for ten fingers through a 1 mm glove but only two fingers through a 5 mm glove, and a millimetre of nitrile attenuates the signal by half. Raise the gain to see through a leather glove and the same gain makes a film of water on the surface look like a finger, because water rejection depends on discriminating a large diffuse patch and that discrimination needs signal-to-noise headroom. This is one knob, not two. Any datasheet claiming excellent glove and water performance simultaneously is describing two firmware builds.

0.198 mm: the pixel pitch sets the viewing distance

One final number, and it is the one that most often goes unexamined. The active area of the 4.3-inch High Brightness 1000nits TFT LCD is 95.04 mm wide across 480 pixels, so the pitch is 0.198 mm and the resolution is about 128 pixels per inch. Angular size at a given distance follows directly: at 400 mm a pixel subtends 1.7 arcminutes, at 600 mm it is 1.14, and at 1 m it is 0.68. Human visual acuity is about one arcminute.

So this is a 600 mm display. That is not a criticism, it is the intended use: a machine panel, an EV charger, a benchtop instrument, a control station where the operator sits at arm's length. Mount it where a 2.4-inch handheld display sits and the individual pixels become visible, which reads as "cheap" to an operator even when the content is well designed. Mount it at 1 m and the resolution is wasted. There is a design consequence either way, and it costs nothing to fix in the concept phase.

The same pitch determines how much UI fits. A 9 mm target is 45 pixels, and allowing a 2 mm gutter that is a grid of eight across and four down: thirty-two targets, comfortably more than double what a 240 x 320 panel offers. At 12 mm it is eighteen, at 15 mm fifteen, at 20 mm eight. The research on bare-thumb accuracy puts the minimum at 9.2 mm for discrete selection, and kiosk guidance for index-finger use recommends 20 mm, so the number you pick is really a statement about gloves and about how many buttons belong on the screen. A legible character needs a cap height of eight or nine pixels, which on 272 rows is 20 to 24 lines of text. If the brief calls for fifteen buttons and three lines of status, this panel has room to spare; if it calls for sixty buttons, the honest answer is that the brief is the problem.

One property that does scale with size is worth a sentence, because it is counterintuitive. Effective contrast in ambient light is a ratio, and ratios do not care about area: a 4.3-inch panel at 1000 cd/m² behind a 4% reflective front stack gives the same effective contrast as a 2.4-inch one at the same luminance. What scales is everything else. The solar load on the front grows with area, so thermal management gets harder while contrast stays the same. The light guide gets longer, so uniformity gets harder. And the operator's eye integrates a larger luminous field, so a screen that is merely bright starts to feel glaring at close range. Brightness and size are separate axes, and the second one is where this panel's design effort belongs.

What this means for the schematic you are about to draw

The honest summary is that the 4.3-inch High Brightness 1000nits TFT LCD is not a harder version of the small panel. It is a different product, and the three numbers that separate them are all on the datasheet. 255 KiB says the memory architecture comes first. 27 MB/s says the bus is always on and the ground return needs engineering. 58 µs says the real-time deadline has moved from the frame to the line, and that your RTOS's worst-case jitter is now a visible artifact.

So sequence the work accordingly. Fix the memory: pick the framebuffer size, the colour depth and the external memory, and confirm the double-buffer case fits with room for the application. Then fix the bus: count the ground pins, place the series resistors, and route the FPC tail before the PCB outline is frozen. Then budget the deadline: measure underrun interrupts, not frame rates. Only after all three does the touch decision belong in the conversation, and it should arrive with the housing drawing rather than after it.

The complete electrical, optical, mechanical and reliability data, including the 40-pin assignment, the AC characteristics with the HSYNC period window, and the quality and inspection section, is in the SFTO430JS-7140BN Datasheet 4.3-inch 1000nits TFT LCD.pdf. If the product needs a capacitive touch panel, a different LED topology for a 12 V or 24 V rail, or a modified FPC, Saef Technology Limited builds those as standard variations on the same cell, and the useful starting point is the framebuffer plan, the enclosure section and the rail voltage rather than a request for a quotation.

The pixels are not the hard part. The memory behind them usually is.