This is twice the required alignment for modern SSE, but it has a clear reason. On both Intel and AMD systems, a 32-byte alignment was ~15% faster than a 16-byte alignment when used with memcpy, which just so happens to be one of the main uses for these frames. Unless the encoder internally keeps its own queue of AVFrames and copies them itself, this will result in a massive speed up across all encoders.
Michael Fabian 'Xaymar' Dirks