Massive multiple-input/multiple-output communication systems are a great solution to satisfy the demand of high-throughput, reliable information transfer, without the need of increasing the required frequency spectrum. In such systems, the base stations are equipped with a very large number of antennas in comparison to the number of users served. Using linear techniques, such as maximum-ratio combining or zero-forcing linear equalization, the effective channels between the base station and the users become almost deterministic, an effect known as channel hardening, even though the actual channel realizations are random. However, to reap these benefits, accurate estimates of the channel coefficients are required, a problem which is very challenging, due to the very large number of these coefficients. Additionally, when pilot-sequence-based techniques are applied, the estimates may suffer from additional interference due to the pilot contamination problem, that arises from the reuse of the limited number of these sequences in the multi-cellular environment. To completely avoid these aforementioned drawbacks, noncoherent detection approaches in massive MIMO systems were proposed. Using differentially-encoded transmit data, and high-performing detection algorithms at the receiver, results that compete with the conventional coherent detection schemes that are typically employed in these systems are achieved. Moreover, only statistical knowledge of the situation, and not the actual conditions, are required. However, a problem still remains. With the hundreds, if not more, radio front-ends, the hardware may become impractical to implement using the state-of-the-art components that are typically designed for single-antenna, or multi-antenna systems that employ only a handful of them. Additionally, to obtain the competitive results, processing is performed on large-dimensional data. This adds to the operational costs, as powerful signal-processing hardware becomes mandatory, when, e.g., latency is critical. Hence, in this dissertation, the design of noncoherent receivers is explored. In the first part, the goal is achieving the best power efficiency possible. This includes improvements in the various steps of the receiver; starting with the antennas, and the radiation pattern they exhibit, the feedback gain in the detection algorithms, the metric utilized to decide which symbols are the most reliable, and decode them, and the combination of differential encoding with error-correction codes to reduce the error rates further. These advanced receiver concepts highlight the potential of noncoherent detection in massive MIMO systems. In the second part of the dissertation, the main focus lies on reducing the algorithmic and hardware complexity of the noncoherent massive MIMO receiver. First, low-complexity alternatives or implementations are adopted for particular computationally-demanding processing tasks, e.g., the SVD for subspace tracking, at the receiver. Next, the structure of the different matrices involved in the detection process are exploited to reduce the numerical complexity from the start. Then, low-resolution analog-to-digital converters are investigated and optimized, to obtain as-good results as the unquantized case. Finally, the special case of one-bit quantization is studied, with the accompanying derivations to obtain a quantization-aware receiver. This entails a solution for acquiring the statistical knowledge needed for detection at the receiver. In contrast to coherent detection employing one-bit converters, it can be proven that at a very high signal-to-noise ratio, it is possible for noncoherent detection to have no error floor. The theoretical insights in this work are supported by numerical results obtained from Monte--Carlo simulations. Additionally, when possible, analytic expressions of the required complexity are derived.