The new generation video codecs typically adopt a sophisticated interpolation filter for sub-pixel motion estimation (ME). For embedded applications, it is crucial to reduce both the memory requirement and the computational complexity for sub-pixel ME. The paper's key observation is that the interpolation filter for motion estimation and the filter for coding do not have to be the same. By adopting a simpler on-the-fly interpolation filter during the ME process and a standard-compliant filter for coding, both memory and complexity can be reduced with very little coding performance degradation. Since Hadamard transformed-SAD (sum of absolute differences) is often used in high quality codecs as a better matching measure, we further show that the ME interpolation filer can be combined with the Hadamard transform for efficient VLSI implementation. Initial results show a very promising performance by the proposed system.