Hyperdimensional computing (HDC) is a lightweight machine-learning paradigm in which data are encoded as high-dimensional hypervectors, enabling error-resilient and hardware-friendly computation. This makes HDC well suited for low-resource embedded systems that must balance performance and energy. However, existing HDC accelerators require full similarity checks between query and class hypervectors, incurring significant computational and energy cost. We propose an HDC algorithm that progressively excludes low-similarity classes by iterating over segmented hypervectors, reducing the number of required comparisons without storing full class representations. Our RTL accelerator reduces execution time by 3.7× and energy consumption by 28.7%, with only a 3.2% accuracy loss.