MCU Hardware

Home > TinyML > MCU HardwareUpdated Mar 2026
memory

Microcontroller Platforms

Arduino Nano 33 BLE Sense

The canonical TinyML teaching board. Packs a Cortex-M4 (64 MHz, 256 KB RAM, 1 MB Flash) with 9-axis IMU, PDM microphone, temperature, humidity, pressure, and colour/proximity sensors. Supported out-of-the-box by TFLite Micro and Edge Impulse.

Key Features
  • CPU: nRF52840 Cortex-M4 @ 64 MHz
  • RAM: 256 KB | Flash: 1 MB
  • Built-in: IMU, mic, env sensors, colour sensor
  • BLE 5.0 radio
  • Edge Impulse first-class support
Similar Technologies
Arduino Nicla Sense MESparkFun EdgeSeeed XIAO BLE
ESP32-S3

Espressif's AI-oriented MCU: dual-core Xtensa LX7 at 240 MHz with vector extensions for INT8 ML, up to 8 MB PSRAM, Wi-Fi + BLE, and a USB OTG. Excellent price-performance for camera + audio TinyML at ~$3/unit.

Key Features
  • CPU: dual Xtensa LX7 @ 240 MHz
  • RAM: 512 KB SRAM + up to 8 MB PSRAM
  • Flash: up to 16 MB
  • Wi-Fi 802.11 b/g/n + BLE 5.0
  • AI vector instructions (INT8 SIMD)
Similar Technologies
ESP32-P4 (higher perf)ESP32-C6Nordic nRF9161
STM32 Cortex-M Family

STMicroelectronics' broad MCU range from Cortex-M0 (ultra-low-power) to M7 (1 MB RAM, 480 MHz). STM32Cube.AI and STEdgeAI tools convert ONNX/TFLite models and optimise for STM32 targets with cycle-accurate profiling.

Key Features
  • Range: M0+ (16 KB RAM) to M7 (1 MB RAM)
  • STM32Cube.AI (free model converter)
  • X-CUBE-AI integration for inference
  • Large ecosystem: HAL, FreeRTOS, Zephyr
  • STM32N6 adds dedicated NPU
Similar Technologies
Nordic nRF seriesNXP i.MX RTRenesas RA series
Nordic nRF52840

ARM Cortex-M4 with BLE 5.0/Bluetooth Mesh/Thread/Zigbee. Industry standard for wearables and IoT. 256 KB RAM supports modest neural nets for keyword spotting and gesture recognition with duty-cycled BLE data transmission.

Key Features
  • CPU: Cortex-M4 @ 64 MHz with FPU
  • RAM: 256 KB | Flash: 1 MB
  • BLE 5.0, Bluetooth Mesh, Thread, Zigbee
  • Ultra-low power modes (1.5 µA sleep)
  • Arm TrustZone security
Similar Technologies
nRF5340 (dual-core)nRF9160 (LTE-M)Silicon Labs EFR32
table_chart

TinyML Hardware Comparison

Board / MCUCoreFlashRAMClockTinyML FrameworksBest For
Arduino Nano 33 BLE SenseCortex-M41 MB256 KB64 MHzTFLite Micro, Edge ImpulsePrototyping, education
ESP32-S3Xtensa LX7 (dual)Up to 16 MB512 KB + 8 MB PSRAM240 MHzTFLite Micro, Edge ImpulseWi-Fi/BT, camera, audio
STM32H7 seriesCortex-M72 MB1 MB480 MHzTFLite Micro, STM32Cube.AIHigh-perf industrial ML
Nordic nRF52840Cortex-M41 MB256 KB64 MHzTFLite Micro, Edge ImpulseBLE wearables, keyword detect
Raspberry Pi RP2040Cortex-M0+ (dual)External (2–16 MB)264 KB133 MHzTFLite Micro, microTVMCost-sensitive prototyping
MAX78000Cortex-M4 + CNN HW512 KB128 KB100 MHzMaxim AI SDKUltra-low power vision
GreenWaves GAP9RISC-V (10-core)External1.7 MB L2370 MHzGAPflow / ONNXAudio & sensor fusion
sensors

Common Sensors for TinyML

IMU (Accelerometer / Gyroscope)

Most common TinyML sensor. 3-axis accelerometer + gyroscope provides motion data for gesture recognition, fall detection, activity classification, and vibration-based anomaly detection. Low power: typically 10–300 µA.

Key Features
  • Common: LSM6DS3, MPU6050, ICM-42688
  • Output: 6-DoF at up to 6.4 kHz
  • I2C / SPI interface
  • 10–300 µA active current
  • Use case: gesture, activity, vibration
PDM / I2S Microphone

Pulse-Density Modulation microphones (e.g., MP34DT05) enable keyword spotting and sound classification on MCUs. Low-cost, low-power, and directly digitally interfaced — no ADC needed.

Key Features
  • Common: MP34DT05, SPH0641LM4H
  • 16-bit audio at 16–32 kHz
  • PDM or I2S output
  • Active current: ~650 µA
  • Use case: keyword spotting, sound events
Image Sensors (QVGA / QQVGA)

Low-resolution CMOS sensors (e.g., OV2640, HM01B0) enable person detection and simple vision tasks on MCUs. The HM01B0 QVGA sensor consumes under 2 mW — enabling always-on vision on battery devices.

Key Features
  • HM01B0: 320×240, <2 mW always-on
  • OV2640: 2 MP, widely available
  • DVCP / SPI / parallel interface
  • QQVGA (160×120) fits in 256 KB RAM
  • Use case: person detect, face detect
precision_manufacturing

Dedicated AI Accelerator MCUs

DeviceVendorAcceleratorPeak TOPSPowerNotes
MAX78000Analog Devices (Maxim)CNN hardware engine~0.5 TOPS<1 mW inferenceUltra-low power, fixed CNN topology
STM32N6 (NPU)STMicroelectronicsNeural Processing Unit600 GOPs~10 mWSTM32 ecosystem, STEdgeAI
Syntiant NDP120SyntiantDeep Learning ProcessorCustom<140 µWAlways-on audio, keyword spotting
Ambiq Apollo510AmbiqNeural Spot (SPOT)Custom<1 mWWearable health AI
GreenWaves GAP9GreenWaves Technologies10-core RISC-V cluster~0.25 TOPS~20 mWAudio + sensor fusion