The Ananse Reports: The Boards and the Measurement Rig
The hardware companion to Report No. 1 documents the boards on the bench, the instruments used to measure them, and the reason the series begins at the TinyML floor.
The series is built around measured results as opposed to estimates. This is the hardware behind that claim: three microcontroller development boards, a power and telemetry rig, and the wiring that makes the measurement boundary visible.
Methodology hardware
Measurement rig
Select a frame, or open the image to inspect markings.
Full rig
Bench-wide view of the rig used to capture board-level energy and telemetry.
Methodology hardware
Measurement rig
Why start with microcontrollers
The Ananse Reports begin with microcontrollers and proceed through single-board computers, accelerators, laptops, and cloud GPUs. This first report focuses on microcontrollers because they expose the constraints that define TinyML. Memory is scarce, and small changes in power draw can determine whether a deployment is practical.
These three boards are not meant to represent every microcontroller. That said, they give the first report three useful comparison points: a conventional embedded baseline, a faster connected MCU, and a sensor-rich TinyML endpoint.
The three boards
STM32 NUCLEO-F401RE
The NUCLEO-F401RE serves as the conventional embedded baseline. It has no onboard radio, microphone array, or environmental sensor suite. Built around an STM32F401RE, its Arm Cortex-M4-class microcontroller runs at 84 MHz with 512 KB of flash and 96 KB of SRAM, and the board includes a built-in ST-LINK programmer/debugger. The question is straightforward: what does Cortex-M4F inference cost on a conventional embedded development board?
Spec check: ST's product page identifies the NUCLEO-F401RE as an STM32 Nucleo-64 board with Arduino Uno V3 and ST Morpho expansion plus integrated ST-LINK. The photographs below document the board used in the experiment, including a scale reference.
Microcontroller under test
STM32 NUCLEO-F401RE
Select a frame, or open the image to inspect markings.
Perspective
Default board view for the classic embedded development board.
Microcontroller under test
STM32 NUCLEO-F401RE
ESP32-S3 DevKitC
The ESP32-S3 DevKitC is the connected MCU in the lineup. This module family combines a 240 MHz dual-core Xtensa processor with Wi-Fi and Bluetooth Low Energy. Many variants add external PSRAM, which eases the memory pressure typical of smaller microcontroller boards. Connected-device teams need to know whether a faster, radio-capable MCU uses less energy per inference or simply finishes sooner.
Spec and price check: Espressif documents the ESP32-S3-DevKitC-1 as an entry-level development board built around ESP32-S3-WROOM modules with Wi-Fi and Bluetooth Low Energy. The bench unit uses the ESP32-S3-WROOM-1-N8R8 module, for which Espressif specifies 8 MB of flash and 8 MB of PSRAM. On August 16, 2026, DigiKey listed this exact variant at $15 for one unit. The photographs below document the board used in the experiment, including scale references.
Microcontroller under test
ESP32-S3 DevKitC
Select a frame, or open the image to inspect markings.
Perspective
Default board view for the connected MCU in the Report No. 1 matrix.
Microcontroller under test
ESP32-S3 DevKitC
Arduino Nano 33 BLE Sense Rev2
The Nano 33 BLE Sense Rev2 is the sensor-rich endpoint. It is the smallest board in the lineup and the one most explicitly aimed at TinyML demos: nRF52840 at 64 MHz, 1 MB Flash, 256 KB SRAM, Bluetooth Low Energy, microphone, inertial sensing, and environmental sensing on a compact Nano-format board. For the Nano 33, I wanted to know what inference would cost on a board that already resembles deployable hardware.
Spec and price check: On August 16, 2026, Arduino's store listed the Rev2 at $38.70. Its published specifications identify the nRF52840, 3.3 V operation, a 64 MHz clock, 1 MB of flash, and 256 KB of SRAM. The photographs below document the board used in the experiment, including tape measurements of its length and width.
Microcontroller under test
Arduino Nano 33 BLE Sense Rev2
Select a frame, or open the image to inspect markings.
Perspective
Default board view for the lowest-energy MCU in Report No. 1.
Microcontroller under test
Arduino Nano 33 BLE Sense Rev2
How the measurements are taken
The measurement rig records three things.
The first is wall-side power: the FNB58 measures what comes in over USB before the board-level regulators and support circuits. The second is rail-side current: INA219 sensors sit in the board-side path, with an MCP2221A bridge carrying I2C telemetry back to the host. The third is ambient context: a BME280 records temperature, humidity, and pressure around the bench so each run has environmental context attached to it.
Those three streams are what make the measurement checkable. A result can be called measured while leaving out where the measurement was taken. This rig makes the boundary explicit: wall side, rail side, ambient conditions, and run IDs that tie each result back to signal-bench.
What comes next
Report No. 1 measures these three boards on three TinyML workloads. Later reports will introduce the next hardware tiers as the series moves beyond microcontrollers. This post documents the starting point and the bench behind those first measurements.
Réponse lecteur
Questions, corrections et objections bienvenues
Si vous avez une correction, une objection méthodologique ou une question de suivi, envoyez-la avec le contexte de l'article afin que je sache exactement à quoi vous répondez.
Your own weekly digest, built around your topics
Callabash curates AI and dev tool research for your role, industry, and stack — and adds the context you need to act on it. Reports like this arrive in it.