THIRD QUARTER 2023 IEEE CIRCUITS AND SYSTEMS MAGAZINE 11 Table 2. Specification and performance comparison of recent progress on TinyML research targeting microcontrollers. CMSIS-NN arXiv'18 [41] On-Device Training or Inference Measured Device Dataset Model Input Resolution Width Multiplier Data Bitwidth Latency Peak Memory Flash Usage Energy Consumption Top-1 Accuracy Inference MCU (STM32H743) ImageNet MobileNetV1 192 0.5 INT8 510 ms1 < 1 MB2 1.4 MB1 135 mJ1 59.5%1 MCUNetV1 NeurIPS'20 [8] On-Device Training or Inference Measured Device Dataset Model Input Resolution Inference MCU (STM32H743) ImageNet MCUNet 160 X-Cube-AI [42] Inference MCU (STM32H743) ImageNet MobileNetV1 192 0.5 INT8 437 ms1 < 1 MB2 1.4 MB1 115 mJ1 59.5%1 TF-Lite Micro MLSys'21 [47] Inference MCU (STM32F746) ImageNet MobileNetV2 64 MicroTVM [43] Inference MCU (STM32F746) CIFAR-10 SmallCifar 32 1.0 INT8 157 ms3 144 KB3 < 1 MB2 - - MicroNets MLSys'21 [48] Inference MCU (STM32F746) VWW MicroNet- VWW-1 160 Liberis et al. MLSys'20 [44] Inference MCU (STM32F767) VWW MobileNetV1 96 0.25 INT8 1325 ms 55 KB < 2 MB2 735 mJ ∼76% MCUNetV2 NeurIPS'21 [9] Inference MCU (STM32H743) ImageNet MCUNet 224 Rusci et al. MLSys'20 [45] Inference MCU (STM32H743) ImageNet MobileNetV1 224 0.75 Mixed 1860 ms < 512 KB2 2 MB 491 mJ4 68.2% TinyOps CVPRW'22 [49] Inference MCU (STM32F746) ImageNet MNASNet 96 CMix-NN TCAS'20 [46] Inference MCU (STM32H743) ImageNet MobileNetV1 192 0.5 INT8 677 ms < 512 KB2 1.4 MB 179 mJ4 62.9% TinyMaix [50] Inference MCU (STM32H750) VWW MobileNetV1 96 (Continued)