This paper investigates what kinds of choices exist in determining the voting structure for cascaded triple modular redundant (TMR) modules, how to identify efficient voting structures, and the effects of voting structures on the overall system reliability. While the classic single-voter and three-voter architectures have been used for about fifty years, the paper shows that there are more practically useful voting structures which provide efficient trade-offs between hardware overhead and system reliability. Specifically, a single-voter architecture with the voter fanout of one is found to be more reliable than the classic single-voter architecture for a reasonably wide range of component reliability.