TinyML on Microcontrollers
Designing and compressing neural networks and inference engines so that vision, audio and sensor models run on microcontrollers with kilobytes of memory.
Tier Grades
Tier A: Frontier AI labs
feasible
The compute is tiny; observed actors are chip vendors and university labs (MIT, Arm, STMicroelectronics) rather than frontier labs.
Tier B: Indonesian AI labs
feasible
Neural architecture search for microcontrollers (MCUNet) needs modest compute; Indonesian work is not in the sources.
Tier C: Institution units such as ITB
feasible
MCUNet came from a 6-author university team.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Vendor model zoos and MLPerf Tiny reference models deploy on low-cost boards for field proofs of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeMCUNet: over 70% ImageNet top-1 on an off-the-shelf microcontroller, with 3.5 times less SRAM and 5.7 times less Flash than quantised MobileNetV2 and ResNet-18.
- DataMLPerf Tiny: four benchmarks (keyword spotting, visual wake words, image classification, anomaly detection), built by more than 50 organisations.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 6 authors on MCUNet; 22 on MLPerf Tiny.
- Time horizonOne to two years.
- Shelf-life signalsMicrocontrollers with neural processing units (STM32N6) widening model sizes, MLPerf Tiny as the shared yardstick for accuracy, latency and energy
- Pipeline roleTier A and chip vendors supply engines and model zoos; Tier C designs models for local sensors; Tier D deploys them on devices in agriculture, factory and building proofs of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
MCUNetTinyMLexemplar
MLPerf Tiny MobileNet (VWW)TinyMLexemplar
FOMO (Faster Objects, More Objects)TinyMLexemplar
ST YOLO detectorsTinyMLexemplar
BC-ResNet (Broadcasted Residual Network)TinyMLproduced byTENet (Temporal Efficient Neural Network)TinyMLproduced by
EEG-TCNetTinyMLproduced byDTLN (Dual-signal Transformation LSTM Network)TinyMLproduced by
DS-CNNTinyMLproduced byHello Edge keyword spotting modelsTinyMLproduced byHello Edge KWS CNN (small, medium, large)TinyMLproduced byHello Edge KWS DNN (small, medium, large)TinyMLproduced byMicroNetTinyMLproduced bymcunet-in4TinyMLproduced bymcunet-vww2TinyMLproduced byMLPerf Tiny 1D DS-CNN (streaming wakeword)TinyMLproduced byMLPerf Tiny Dense AE (ToyCar)TinyMLproduced byMLPerf Tiny DS-CNN (KWS)TinyMLproduced byMLPerf Tiny ResNet (CIFAR-10)TinyMLproduced byMobileNet v2TinyMLproduced byFOMO-AD (Visual Anomaly Detection)TinyMLproduced by
Micro Speech (TFLM keyword model)TinyMLproduced by
Microspeech LSTMTinyMLproduced byRNNoiseTinyMLproduced by+17
SESRTinyMLproduced by
Fingerprint Signature GeneratorTinyMLproduced byKeyword Spotting NumbersTinyMLproduced byKeyword Spotting On/Off v3TinyMLproduced bySilicon Labs MLTK keyword spotting modelsTinyMLproduced bySSD MobileNet v1TinyMLproduced byST arc fault detection models (st_conv, st_dense)TinyMLproduced byST CNN2D hand posture (Time-of-Flight)TinyMLproduced byST human activity recognition models (st_gmp, st_ign)TinyMLproduced byST MiniResNet v1 and v2 audio event detectionTinyMLproduced byST MNIST v1TinyMLproduced byST YOLO LC v1TinyMLproduced byST YOLOD v2 (Tiny, Milli)TinyMLproduced byST YOLOX-nanoTinyMLproduced bySTFT-TCNN speech enhancementTinyMLproduced byWav2letterTinyMLproduced byYOLO v3 TinyTinyMLproduced by
Fingerprint Signature GeneratorTinyMLproduced byKeyword Spotting NumbersTinyMLproduced byKeyword Spotting On/Off v3TinyMLproduced bySilicon Labs MLTK keyword spotting modelsTinyMLproduced bySSD MobileNet v1TinyMLproduced byST arc fault detection models (st_conv, st_dense)TinyMLproduced byST CNN2D hand posture (Time-of-Flight)TinyMLproduced byST human activity recognition models (st_gmp, st_ign)TinyMLproduced byST MiniResNet v1 and v2 audio event detectionTinyMLproduced byST MNIST v1TinyMLproduced byST YOLO LC v1TinyMLproduced byST YOLOD v2 (Tiny, Milli)TinyMLproduced byST YOLOX-nanoTinyMLproduced bySTFT-TCNN speech enhancementTinyMLproduced byWav2letterTinyMLproduced byYOLO v3 TinyTinyMLproduced byLinks
R&D TopicTinyML on Microcontrollers
ModelsMCUNetMIT Han Lab, TinyMLMLPerf Tiny MobileNet (VWW)MLCommons, TinyMLFOMO (Faster Objects, More Objects)Edge Impulse, TinyMLST YOLO detectorsSTMicroelectronics, TinyML+37 more
MCUNetMIT Han Lab, TinyMLMLPerf Tiny MobileNet (VWW)MLCommons, TinyMLFOMO (Faster Objects, More Objects)Edge Impulse, TinyMLST YOLO detectorsSTMicroelectronics, TinyML+37 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 41
ST YOLO detectorsSTMicroelectronics, TinyMLexemplarHello Edge keyword spotting modelsArm, TinyMLproduced bySilicon Labs MLTK keyword spotting modelsSilicon Labs, TinyMLproduced byFOMO-AD (Visual Anomaly Detection)Edge Impulse, TinyMLproduced byHello Edge KWS CNN (small, medium, large)Arm, TinyMLproduced byHello Edge KWS DNN (small, medium, large)Arm, TinyMLproduced byKeyword Spotting NumbersSilicon Labs, TinyMLproduced byKeyword Spotting On/Off v3Silicon Labs, TinyMLproduced byMCUNetMIT Han Lab, TinyMLexemplarST YOLO LC v1STMicroelectronics, TinyMLproduced byST YOLOD v2 (Tiny, Milli)STMicroelectronics, TinyMLproduced byST YOLOX-nanoSTMicroelectronics, TinyMLproduced by+29
BC-ResNet (Broadcasted Residual Network)Qualcomm, TinyMLproduced byDS-CNNArm, TinyMLproduced byFOMO (Faster Objects, More Objects)Edge Impulse, TinyMLexemplarmcunet-in4MIT Han Lab, TinyMLproduced bymcunet-vww2MIT Han Lab, TinyMLproduced byMicro Speech (TFLM keyword model)Google, TinyMLproduced byMicrospeech LSTMNXP Semiconductors, TinyMLproduced byMLPerf Tiny 1D DS-CNN (streaming wakeword)MLCommons, TinyMLproduced byMLPerf Tiny Dense AE (ToyCar)MLCommons, TinyMLproduced byMLPerf Tiny DS-CNN (KWS)MLCommons, TinyMLproduced byMobileNet v2TinyMLproduced byST arc fault detection models (st_conv, st_dense)STMicroelectronics, TinyMLproduced byST MiniResNet v1 and v2 audio event detectionSTMicroelectronics, TinyMLproduced bySTFT-TCNN speech enhancementSTMicroelectronics, TinyMLproduced byDTLN (Dual-signal Transformation LSTM Network)TinyMLproduced byEEG-TCNetETH Zurich, TinyMLproduced byFingerprint Signature GeneratorSilicon Labs, TinyMLproduced byMicroNetTinyMLproduced byMLPerf Tiny MobileNet (VWW)MLCommons, TinyMLexemplarMLPerf Tiny ResNet (CIFAR-10)MLCommons, TinyMLproduced byRNNoiseTinyMLproduced bySESRTinyMLproduced bySSD MobileNet v1TinyMLproduced byST CNN2D hand posture (Time-of-Flight)STMicroelectronics, TinyMLproduced byST human activity recognition models (st_gmp, st_ign)STMicroelectronics, TinyMLproduced byST MNIST v1STMicroelectronics, TinyMLproduced byTENet (Temporal Efficient Neural Network)TinyMLproduced byWav2letterTinyMLproduced byYOLO v3 TinyTinyMLproduced byPapers
- Hello Edge: Keyword Spotting on MicrocontrollersYundong Zhang et al., arXiv preprint, 2017 founding
- MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep LearningJi Lin et al., arXiv preprint, 2021 state of the art
- MLPerf Tiny BenchmarkColby Banbury et al., arXiv preprint, 2021 state of the art
- MCUNet: Tiny Deep Learning on IoT DevicesJi Lin et al., NeurIPS 2020 (per arXiv comment), 2020 state of the art
- TensorFlow Lite Micro: Embedded Machine Learning on TinyML SystemsRobert David et al., arXiv preprint, 2020 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2007.10319 https://arxiv.org/abs/2007.10319
- arxiv.org/abs/2106.07597 https://arxiv.org/abs/2106.07597
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "TinyML on Microcontrollers", https://aurion.semai.id/topics/tinyml-on-microcontrollers/, records of October 2026.