Hyperdimensional computing is a brain-inspired computing paradigm that performs highly accurate classifications for biomedical applications by operating on pseudo-random hypervectors. However, the energy consumption of existing hyperdimensional computing processors is dominated by the memory storage of hypervectors, which grow linearly with the number of sensor channels. In this work, the memory is replaced with a very light-weight cellular automaton for on-the-fly hypervector generation. Vector folding is explored in conjunction to maximize energy efficiency of many-channeled classification tasks, demonstrated through multiple experiments on an emotion recognition dataset. The proposed architecture achieves 39.4 nJ/prediction in post-layout simulation for classification of >200 channels and 3 physiological sensor modalities; a 4.8x improvement in energy efficiency, 9.6x per channel, over the state-of-the-art hyperdimensional computing processor.