MCU Selector
Five questions, then three defensible candidates with the arithmetic shown - including what each one will cost you later. A distributor's filter returns four thousand matching parts; the decision is never made on the specs it filters on.
Shortlist
Ranked by the scoring described at the bottom of this page. Parts that cannot do the compute class you asked for are removed rather than ranked low.
nRF52832
Best fit- Core
- Cortex-M4F @ 64 MHz
- Memory
- 512 KB flash · 64 KB RAM
- Unit cost · sleep
- $2.40-$4.20 · 1.5 µA
The BLE workhorse. A decade of field history, the best power numbers in the category, and pre-certified modules from half a dozen vendors.
- BLE is on the die.
- Sleeps at about 1.5 µA, which is coin-cell territory in a duty-cycled design.
- Pre-certified modules exist, which is the single biggest schedule saving available on a wireless product.
- Documentation, toolchain and community are mature enough that the first month is spent on your product.
- More capability than the stated requirement, which you pay for in cost, power and board complexity.
What will bite you: 64 KB of RAM gets tight once a soft device, a filesystem and an over-the-air update buffer are all resident. If the product will grow, start on the 52840 instead of migrating later.
nRF52840
- Core
- Cortex-M4F @ 64 MHz
- Memory
- 1 MB flash · 256 KB RAM
- Unit cost · sleep
- $3.80-$6.50 · 0.4 µA
The safe default for anything battery powered with BLE. 256 KB of RAM leaves room for a real application plus updates, 802.15.4 means Thread or Matter later without a new chip, and system-off is comfortably under a microamp.
- BLE is on the die.
- Sleeps at about 0.4 µA, which is coin-cell territory in a duty-cycled design.
- Pre-certified modules exist, which is the single biggest schedule saving available on a wireless product.
- Documentation, toolchain and community are mature enough that the first month is spent on your product.
- More capability than the stated requirement, which you pay for in cost, power and board complexity.
What will bite you: Two to three times the price of an ESP32-C3, and no Wi-Fi. If the product needs a connection to a cloud rather than to a phone, this is the wrong half of the tree.
nRF5340
- Core
- dual Cortex-M33 + TrustZone @ 128 MHz
- Memory
- 1 MB flash · 512 KB RAM
- Unit cost · sleep
- $6.50-$10.00 · 1.2 µA
Two cores, so the radio stack stops stealing your application's timing, plus TrustZone, secure boot and Bluetooth direction finding if indoor positioning is on the roadmap.
- BLE is on the die.
- Sleeps at about 1.2 µA, which is coin-cell territory in a duty-cycled design.
- Pre-certified modules exist, which is the single biggest schedule saving available on a wireless product.
- Documentation, toolchain and community are mature enough that the first month is spent on your product.
- More capability than the stated requirement, which you pay for in cost, power and board complexity.
What will bite you: Dual core means a dual-image build and a dual-image update path, and a steeper first month. Overkill unless you need the isolation, the security posture or the angle-of-arrival feature.
Everything that qualified
The whole field for these answers, so you can see what came close. 2 of the 16 parts in the list were ruled out entirely - they either cannot do the compute class you asked for, or they fail the power requirement outright.
| Fit | Part | Core | Radio | Sleep | Unit cost |
|---|---|---|---|---|---|
| 85% | nRF52832 | Cortex-M4F | ble | 1.5 µA | $2.40-$4.20 |
| 84% | nRF52840 | Cortex-M4F | ble, thread | 0.4 µA | $3.80-$6.50 |
| 83% | nRF5340 | dual Cortex-M33 + TrustZone | ble, thread | 1.2 µA | $6.50-$10.00 |
| 64% | STM32WLE5 | Cortex-M4 + sub-GHz radio | lora | 1.5 µA | $4.00-$7.00 |
| 58% | STM32G0 | Cortex-M0+ | none | 2 µA | $1.20-$3.50 |
| 52% | CH32V003 | RISC-V QingKe V2A | none | 10 µA | $0.10-$0.25 |
| 51% | STM32U5 | Cortex-M33 + TrustZone | none | 1.5 µA | $4.50-$9.00 |
| 43% | ESP32-C3 | RISC-V single core | wifi, ble | 5 µA | $1.40-$2.60 |
| 42% | ESP32-H2 | RISC-V single core | thread, ble | 7 µA | $1.50-$2.80 |
| 37% | ESP32-S3 | dual Xtensa LX7 with vector extensions | wifi, ble | 7 µA | $2.40-$4.50 |
| 25% | Raspberry Pi CM5 | quad Cortex-A76, embedded Linux | wifi, ble | 100000 µA | $45.00-$95.00 |
| 13% | nRF9160 | Cortex-M33 + LTE modem + GNSS | cellular | 2.7 µA | $11.00-$18.00 |
| 10% | RP2350 | dual Cortex-M33 or RISC-V | none | 180 µA | $1.10-$1.80 |
| 2% | STM32H7 | Cortex-M7 (dual-core variants) | none | 600 µA | $9.00-$20.00 |
How the scoring works
Five terms, added up, out of a possible 141. Connectivity is worth up to 40 points, because a radio that is on the die rather than beside it removes a chip, an antenna match and an entire certification path. Compute class is worth up to 25, with a penalty for overshooting rather than a reward: a part that is two classes too big costs you money, power and board area forever. The power source is worth up to 25, and failing it costs 35, because a part that cannot meet the battery requirement is not a compromise, it is the wrong part. Unit cost is worth up to 25 if you said every cent counts, roughly half that if you said balanced, and nothing if you said capability first. The last term, up to 26, is the trade between a ship date and a bill of materials: pre-certified modules and a mature toolchain if you picked the first, a straight ecosystem-maturity score if you picked the second.
Parts that cannot do the compute class you asked for are removed entirely instead of ranked low, which is why a Linux answer makes the microcontrollers disappear. Nothing else is filtered: a part with no radio still appears for a wireless requirement, scored on the assumption that you would add a companion module, because that is a legitimate design and sometimes the right one.
What this cannot decide for you
Three things dominate real selections and none of them are in the model. Availability: a part with a fifty-two week lead time is not a candidate whatever it scores, so check stock at two distributors before you commit a schematic. Who is writing the firmware: a team that has shipped three Nordic products will beat a theoretically better part they have never used, and that is a rational reason to pick the familiar one. And the second product: if the roadmap says the next device needs Wi-Fi, choosing a family that spans both saves a rewrite, which is the one case where deliberately overshooting is correct.
Prices here are bands for the low thousands of pieces and they move. Sleep currents are what a careful design achieves with memory retained, not the best figure in the datasheet's summary table. Treat both as a starting point for a conversation with a distributor, and measure the sleep current on your own board, because the system sleeps at a much higher current than the MCU does once regulators, sensors and pull-ups are counted.
When the choice is load-bearing
Choosing a microcontroller is the first irreversible decision in a hardware project. It fixes your radio, your toolchain, your certification path, your power budget and, in practice, who can maintain the firmware in three years. We make that call for founders most weeks, and we have made it wrong often enough to know where it hurts: how we work with hardware startups, ESP32 vs STM32 vs Arduino for production on the two families this page keeps returning, and a free 30-minute technical call if you want the shortlist argued with rather than agreed with.