Edge AI / TinyML Implementation
TinyML is ML on microcontrollers with RAM in kilobytes and power consumption in milliwatts. This is not simplification—it's a separate engineering discipline with unique constraints and approaches.
TinyML Hardware Spectrum
Tier 1 (Microcontrollers):
- Cortex-M4/M7: STM32H7, nRF52840. RAM 256 KB – 1 MB. Consumption 1–100 mW
- Xtensa LX7: ESP32-S3 with vector instructions. RAM 512 KB. ~240 MHz
Tier 2 (AI-enabled MCU):
- MAX78000/MAX78002 (Maxim/Analog Devices): built-in CNN accelerator, 0.5 mW inference
- STM32N6: built-in NPU 600 GOPS — breakthrough for MCU class
- Arduino Nicla Vision: OV2640 + Cortex-M7 for edge vision
Tier 3 (Edge SBC):
- Raspberry Pi 5 + Hailo-8 (26 TOPS for $20)
- BeagleBone AI-64
Frameworks and Tools
TFLite Micro: Google, most mature. Portable C++, ~100 KB footprint, supports all MCU.
Edge Impulse Studio: cloud platform for training + deploy to MCU. Drag-and-drop for prototyping.
ONNX Runtime for MCU: newer, growing ecosystem.
ExecuTorch (PyTorch): Meta's embedded ML runtime, ARM Cortex-M support.
Typical Workflow
Data Collection → Feature Engineering (no space on MCU for raw data) → Model Design (NAS for specific resource constraints) → Training → Post-Training Quantization (INT8/INT4) → Deployment → Validation.