Large language models (LLMs) have taken the center stage of the artificial intelligence landscape. As they are pervasive, there is a need to evaluate their dependable operation when soft errors appear in the memories. In this chapter, the dependability of LLMs as measure of correct operation in the presence of errors is evaluated based on simulation with fault injection. Two typical error types are considered: errors caused when using approximate memory and soft errors induced by radiation. The impact of errors is evaluated on three main storage components of LLMs, including quantized weights, Key-Value (KV) cache, and text embeddings. For each of them, multiple LLM models are studied, and the effect of errors is evaluated by considering different factors, including bit position, bit-width, layer, inference stage, and so on. In general, the simulation results reveal that only the errors on some significant bits can cause significant performance degradation, and most errors on other bits can be tolerated by the LLMs. Therefore, if the significant bits can be well protected, it is feasible to deploy LLMs on edge devices with approximate memories for power saving purposes, and it is also possible to withstand radiation-induced soft errors in the memory. This chapter provides valuable references for the design of efficient protection schemes for the main parameters and variables in LLMs and their memory usage.