Table 3. Performance comparison of various tiny models and inference frameworks on STM32H743, which runs at 480 MHz with the resource constraint of 512 KB peak memory and 2 MB storage. CMSIS-NN arXiv'18 [41] X-Cube-AI [42] Peak Memory Storage usage Peak Memory Storage usage Peak Memory Storage usage Peak Memory Storage usage Peak Memory Storage usage Peak Memory Storage usage Peak Memory Storage usage 53 ms 163 KB 32 ms 88 KB TinyEngine NeurIPS'20 [8] 27 ms 59 KB TF-Lite Micro MLSys'21 [47] Dataset: VWW; Model: mcunet-vww0; Input Resolution: 64; Width Multiplier: N/A; Top-1 Accuracy: 87.3% Latency 587 ms 163 KB 646 KB 97 ms 220 KB 736 KB 478 ms 390 KB 463 KB 57 ms 113 KB 453 KB 51 ms 92 KB 627 KB Dataset: VWW; Model: mcunet-vww1; Input Resolution: 80; Width Multiplier: N/A; Top-1 Accuracy: 88.9% Latency 1120 ms 220 KB 534 KB 269 ms 201 KB 774 KB 35 ms 69 KB 521 KB 234 ms 174 KB 741 KB 25 ms 49 KB 718 KB Dataset: VWW; Model: mcunet-vww2; Input Resolution: 144; Width Multiplier: N/A; Top-1 Accuracy: 91.8% Latency 5310 ms 385 KB 1034 KB 51 ms 161 KB 1016 KB Dataset: ImageNet; Model: mcunet-in0; Input Resolution: 48; Width Multiplier: N/A; Top-1 Accuracy: 40.4% Latency 596 ms 161 KB 1090 KB 103 ms 219 KB 956 KB 642 ms 469 KB 856 KB 63 ms 106 KB 842 KB 56 ms 96 KB 1072 KB Dataset: ImageNet; Model: mcunet-in1; Input Resolution: 96; Width Multiplier: N/A; Top-1 Accuracy: 49.9% Latency 1227 ms 219 KB 737 KB 351 ms 238 KB 849 KB 414 ms 243 KB 867 KB 727 KB 280 ms 215 KB 830 KB 336 ms 260 KB 835 KB 937 KB Dataset: ImageNet; Model: mcunet-in2; Input Resolution: 160; Width Multiplier: N/A; Top-1 Accuracy: 60.3% Latency 6463 ms 460 KB 1102 KB 770 ms 493 KB 1084 KB Dataset: ImageNet; Model: mcunet-in3; Input Resolution: 176; Width Multiplier: N/A; Top-1 Accuracy: 61.8% Latency 7821 ms 493 KB 1106 KB 1091 KB (Continued) 14 IEEE CIRCUITS AND SYSTEMS MAGAZINE THIRD QUARTER 2023