notesum.ai
Published at December 5MultiTASC++: A Continuously Adaptive Scheduler for Edge-Based Multi-Device Cascade Inference
cs.LG
cs.DC
Released Date: December 5, 2024
Authors: Sokratis Nikolaidis1, Stylianos I. Venieris2, Iakovos S. Venieris1
Aff.: 1National Technical University of Athens, Athens, Greece; 2Samsung AI Center, Cambridge, UK

| Model | Location | Device | Clock Rate | Accuracy | Latency | FLOPs | #Params |
|---|---|---|---|---|---|---|---|
| MobileNetV2 [13] | Low-end | Sony Xperia C5 | 1.69 GHz | 71.85% | 31 ms | 0.6 B | 3.5 M |
| EfficientNetLite0 [14] | Mid-tier | Samsung A71 | 2.20 GHz | 75.02% | 43 ms | 0.8 B | 4.7 M |
| EfficientNetB0 [14] | High-end | Samsung S20 FE | 2.73 GHz | 77.04% | 33 ms | 0.8 B | 5.3 M |
| MobileViT-x-small [28] | High-end | Google Pixel 7 | 2.85 GHz | 74.64% | 57 ms | 1.1 B | 2.3 M |
| InceptionV3 [29] | Server | Tesla T4 GPU | 585 MHz | 78.29% | 15 ms | 11.4 B | 23.8 M |
| EfficientNetB3 [14] | Server | Tesla T4 GPU | 585 MHz | 81.49% | 25 ms | 3.7 B | 12.2 M |
| DeiT-Base-Distilled [30] | Server | Tesla T4 GPU | 585 MHz | 83.41% | 14 ms | 7.7 B | 86.0 M |