12 IEEE CIRCUITS AND SYSTEMS MAGAZINE THIRD QUARTER 2023 Table 2. Specification and performance comparison of recent progress on TinyML research targeting microcontrollers. Width Multiplier Data Bitwidth N/A INT8 Latency Peak Memory Flash Usage Energy Consumption Top-1 Accuracy On-Device Training or Inference Measured Device Dataset Model Data Bitwidth Latency Peak Memory Flash Usage Energy Consumption Top-1 Accuracy 463 ms 416 KB 1.7 MB - 68.0% UDC NeurIPS'22 [51] Inference N/A (Simulation) ImageNet UDC Mixed - - 1.27 MB - 72.1% 0.35 INT8 296 ms3 211 KB3 < 1 MB2 - - TinyTL NeurIPS'20 [52] Training N/A (Simulation) CIFAR-10 ProxylessNASMobile FP32 - 65 MB - - 96.1% N/A INT8 1133 ms 285 KB 0.8 MB 479 mJ 88.0% TinyOL IJCNN'21 [53] Training MCU (nRF52840) Self-Collected Autoencoder FP32 - < 256 KB2 < 1 MB2 - - N/A INT8 859 ms 434 KB 1.8 MB - 71.8% POET ICML'22 [54] Training MCU (nRF52840) CIFAR-10 ResNet-18 FP32 49 ms 271 KB < 1 MB2 868 mJ 95.5% (Continued) 1.0 INT8 866 ms 397 KB 4.7 MB 546 mJ 64.0% MiniLearn EWSN'22 [55] Training MCU (nRF52840) KWS-subset Customized Mixed 93 ms 196 KB 0.9 MB 1486 mJ 88.5% 0.25 Mixed 64 ms 54 KB 0.2 MB - ∼76% MCUNetV3 NeurIPS'22 [56] Training MCU (STM32F746) VWW MCUNet INT8 546 ms 173 KB 0.7MB - 89.3% 1Measured by CMix-NN paper [46]. 2Speculated by the specification of the corresponding MCU. 3Measured by MCUNet paper [8]. 4In a private email on December 22, 2022, the authors of the papers replied that the energy consumption should be interpreted as mJ instead of μJ in their papers.