Embedded artificial intelligence has emerged as an essential technology to realize next-generation embedded services such as self-driving cars, augmented reality, virtual reality and AI camera. For the embedded AI, one of the key technologies is heterogeneous computing that utilizes all available computing units. However, it requires excessive power consumption because of utilizing the all computing units. Since it is crucial to reduce the excessive power consumption, numerous DVFS related work has been proposed to properly lower their operating frequency. However, the existing related work is not suitable to guarantee the desired time of a neural network. In order to overcome the limitation, we propose a neural network characteristics-aware proactive boost for heterogeneous computing. The proposed approach proactively estimates the lowest operating frequencies of NPU, DSP, GPU, CPU, memory controller and system bus to guarantee the desired time of the neural network. We implement our approach into the Galaxy S22 with Exynos 2200. We show the effectiveness of the proposed approach via extensive experiments.