We evaluate two architectures for fixed length distribution matching. Distribution matching is essential for probabilistic amplitude shaping, a modulation scheme recently proposed for next generation wireless communications. The parallel architecture, called the bit-level distribution matcher (BL-DM) is compared with the symbol-level constant composition distribution matcher (CCDM). Simulation results confirm higher throughput of the BL-DM without performance loss compared to the symbol-level CCDM, when multiple processors can be used. For single core systems, the symbol-level CCDM achieves higher throughput. We also prove that when the output distributions of the BL-DM and the symbol-level CCDM are the same, i.e., they can be represented as a product of binary distributions, the BL-DM achieves lower or equal rate-loss.