Optimal Energy-Delay Tradeoff for Mobile Edge Generation (MEG)

A novel mobile edge generation (MEG) framework is proposed to enable large generative model (LGM) capabilities at the edge, which offers low-latency, power-saving, and languageguided generation on mobile device. Specifically, our framework splits the generation of large-scale content (e.g., high-definition image) into two parts, namely primary and secondary regions. Only the primary region is generated by the LGM at the edge cloud and transmitted via downlink, while the secondary region is generated by the tiny generative model (TinyGM) at the mobile device. By configuring generation splitting ratio between edge and mobile devices, the transmission and computation overheads can be reduced. We formulate a joint generation splitting and mobile power control optimization problem. The formulated problem is a multi-objective optimization programming, which simultaneously minimizes the generation latency and the mobile energy consumption. To explore the performance limits, we first transfer the multi-objective programming into a single-objective programming based on the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\epsilon$</tex>-constraint method. Then, we derive the closed-form Pareto-optimal solution of generation splitting and mobile power control. Thereby, the performance boundary of energy-delay (E-D) tradeoff region is obtained. Furthermore, we also identify the conditions under which the proposed MEG strictly outperforms the fully edge generation (FEG) scheme, and demonstrates that performance gains increase as signal-tonoise ratio (SNR) and mobile generation cost decrease. Numerical results demonstrate the optimal E-D tradeoff of the proposed MEG and verify that it can significantly reduce the latency compared to FEG while achieving satisfactory generation.