UART, start to stop

Change the receiver's speed. The sender never finds out, and neither does your code.

One frame, ten bit times

The line idles high. A falling edge — the start bit — is the receiver's only synchronisation, and everything after it is counted, not measured:

1. one low start bit,
2. eight data bits, least significant first,
3. one high stop bit.

Ten bit times per byte, which is why 9600 baud is 960 characters a second and 115200 is 11,520.

The receiver aims for the middle of each bit — half a bit period after the edge, then one period per bit after that. All the tolerance in the system comes from that half-bit of margin: about 2% of clock error is fine, 5% loses the last bits of the frame first.

Wrong baud is not an error

Nothing in the frame says how fast it was sent. Set the receiver too fast and its ten samples all land inside the start bit; too slow and they drift off the end into the stop bit. Either way it assembles a byte, hands it to your program, and waits for the next one.

Two wires, crossed

TX goes to RX and RX goes to TX. There is no address, no master and no negotiation: two devices, one pair, forever. Ground must be common, and if the two boards run at different voltages the 3.3 V side needs level shifting — a 5 V TX into a 3.3 V RX is the standard way to lose a pin.

#define GPS_RX 16   // ESP32 receives here — wire it to the module's TX
#define GPS_TX 17

void setup() {
  Serial.begin(115200);                          // to the computer
  Serial2.begin(9600, SERIAL_8N1, GPS_RX, GPS_TX);   // to the GPS
}

void loop() {
  while (Serial2.available()) Serial.write(Serial2.read());
}

SERIAL_8N1 is the other half of the agreement — eight data bits, no parity, one stop bit — and it is what almost everything uses. The rare device wanting 7E1 or 8N2 will say so in the first paragraph of its datasheet.

Where it wins

A metre of cable, one device, a stream of data that nobody has to request: GPS modules, fingerprint readers, serial LCDs, cellular modems, and one board talking to another. It is the only protocol here that is comfortable off the board, and with an RS-485 transceiver on each end the same frames go a kilometre.

What it cannot do is scale. No addressing means one device per pair of pins, which is exactly the constraint I2C was invented to remove.

The first thing to check

Baud, then wiring, then ground. In that order, because it is the cheapest to test — and because "serial prints garbage" has one dominant cause and it is a number typed into two places that do not match.