One Pull-Up Pair Per I2C Bus

The most common mistake on an I2C schematic is not a swapped pin or a missing decoupling cap. It is four pull-up pairs on a bus that needs one. Every breakout brought its own, nobody removed any of them, and the bus ends up with a pull-up value the designer never chose.

I2C is open drain. No device on the bus ever drives SDA or SCL high. Devices only pull low, through a transistor to ground, and the line returns high because a resistor pulls it toward VDD. That resistor belongs to the bus, not to the chip.

The parallel math

Put four 4.7k pull-ups in parallel and you get 4700 / 4 = 1175 ohms. Not 4.7k. Every module you add makes the pull-up stronger, and nothing in the schematic looks wrong while it happens.

Now the sink side. The I2C specification says a device must hold its low-level output voltage at or below 0.4 V while sinking 3 mA. Turn that around and you get the minimum sensible pull-up resistance for the whole bus:

Rmin = (VDD - 0.4 V) / 3 mA

At 3.3 V that is 2.9 / 0.003 = 967 ohms. At 5 V it is 4.6 / 0.003 = 1533 ohms.

So four 4.7k pairs on a 3.3 V bus land at 1175 ohms. Above the floor, but with almost no margin left. Add a fifth module and you are at 940 ohms, which is below it. On a 5 V bus, four pairs are already out of spec before you add anything at all.

And 3 mA is what the specification demands, not what every part delivers. Plenty of small sensors, GPIO expanders and microcontroller pads sink considerably less and sit at a higher low-level voltage. Once the low stops getting below the input low threshold, which is 0.3 x VDD, or 0.99 V on a 3.3 V bus, the receiver no longer sees a valid low.

What you observe is never a clean failure. It is a bus that enumerates perfectly on the bench, then NACKs at 40 C, or locks up after an hour, or works fine until the fourth board is plugged in.

The other end of the trade: rise time

Weak pull-ups fail the opposite way. The line rises through an RC formed by the pull-up and the total bus capacitance, which includes pads, traces, cable, and every device's pin capacitance.

Measured the way the standard measures it, from 0.3 to 0.7 x VDD, the rise time is:

t = 0.8473 x R x C

The budgets are 1000 ns at 100 kHz standard mode, 300 ns at 400 kHz fast mode, and 120 ns at 1 MHz. Maximum bus capacitance is 400 pF in standard and fast mode, 550 pF in fast-mode plus.

Work an example. 4.7k with 50 pF of bus capacitance gives 0.8473 x 4700 x 50 pF = 199 ns, comfortable at 400 kHz. The same 4.7k with 100 pF gives 398 ns, which misses the fast-mode budget and only works at 100 kHz. With a 4.7k pull-up, your entire 400 kHz rise budget is spent at about 75 pF.

That is exactly why 4.7k at 3.3 V is a good default for a short on-board bus. It sits between the 967 ohm floor and the capacitance ceiling with room on both sides, for the few tens of picofarads that a few centimeters of trace actually carry. It stops being a good default the moment you add a meter of cable.

Modules that bring their own

Read the schematic of every breakout going on the bus. Most vendors publish them. Values vary a lot: 10k, 4.7k and 2.2k are all common, and mixed values still parallel down to something small.

Then pick one pull-up pair for the entire bus and remove the rest. Most breakouts have either a solder jumper you can cut or a pair of 0402s you can drag off with a hot iron. If a module's resistors are genuinely unreachable, budget deliberately: keep that module's pair, strip every other one, and check the result against the 967 ohm floor.

If you are designing the board rather than wiring modules together, the problem mostly disappears. Put the sensors on your own PCB and place one pull-up pair near the controller.

How MML-01 does it

MML-01 runs a BME280 and a second sensor on one shared I2C bus off the ESP32-S3. There is a single 4.7k pull-up pair on the whole board, placed near the microcontroller, and neither sensor contributes one, because there is no breakout involved. The parts sit on my PCB, not on somebody else's.

That is the whole trick. The pull-up pair is a property of the bus, so it is drawn once, in one place, and it appears exactly once in the BOM.

If a sensor has to live off-board on a header, the cable becomes part of the bus capacitance. Keep it short, keep the clock at 100 kHz, and recompute the rise time before assuming it is fine.

The practical version

Count the pull-ups on your finished schematic. If the number of resistors tied to SDA is anything other than one, fix it before you route.

Pull-ups belong to the bus, not to the device: one pair, chosen on purpose.