This report presents some fundamentals of floating-point (FP) arithmetic and some causes of substantial loss of accuracy and examines the internal FP representations for large-scale minicomputers. Results of test of relative accuracy, which involved executing a set of test programs on several representative systems, are given along with the some conclusions regarding the accuracy inherent in each system. The systems used were the Harris Series 500, VAX 11/780, Prime 550, and IBM 4331. The CDC CYBER 6600 was used to establish a baseline for comparing the results. (Author).