This paper discusses design features in currently available RISC microprocessors that sametimes result in less-than-optimal sustinded performance on large-scale scientific caltulations. Recommendations for future designs are suggested.