Gaussian Pyramid (GP) is one of the most important representations in computer vision. However, the computation of GP is still challenging for real-time applications. In this paper, we propose a novel approach by investigating the extended box filters for an efficient Gaussian approximation. Taking advantages of the cascade configuration, tiny kernels and memory cache, we develop a fast and suitable algorithm for embedded systems, typically smartphones. Experiments with Android NDK show a 5x speed up compared to an optimized CPU-version of the Gaussian smoothing.