The two best-developed computational models of reading aloud, the DRC model of Coltheart and colleagues and the connectionist attractor model of Plaut and colleagues, offer very different views about the degree to which semantic knowledge is involved in lexical processing, and hence make differing predictions about how semantic impairment (as seen, for example, in semantic dementia) will impact on lexical processing in clinical cases. Two cases meeting the criteria for semantic dementia, PC and EM, were given a battery of tests comprising comprehension tasks, a reading task, and a visual word recognition (lexical decision) task. AD tasks used the same target words allowing cross-test and cross-patient comparisons. Both cases showed significant impairment of semantic memory, and word comprehension was found to be related to the word frequency of the target words. PC demonstrated poor reading of irregular words, with a surface dyslexic pattern of reading aloud, and he performed poorly on the visual lexical decision task. His ability to read irregular words was related to their frequency and to his ability to comprehend them. In contrast, his visual lexical decision performance was not reliably influenced by his comprehension of the same words or by their frequency. EM demonstrated essentially perfect reading aloud of irregular words and essentially perfect visual lexical decision, despite her severe semantic impairment. The pattern of performance shown by EM is consistent with the DRC model of reading, but inconsistent with the connectionist attractor model and with the view, associated with that model, that orthographic and phonological processes cannot remain intact when semantic representations are degraded.