Five major DFT algorithms were evaluated on seven different computers. The relative performances of these algorithms were related to the architecture of each computer by finding a relationship between the execution time and the instruction counts. The relative performance of these algorithms on other Computers is predicted, based on the knowledge of the computer architecture. On certain implementations, data transfers are more important than floating-point additions and multiplications when comparing DFT algorithms. On the average, data transfers account for a greater percentage of the execution time than floating-point operations. W TNTRODUCTION E have been unable to find in the literature an evaluation of the execution time of the basic DFT algorithms as a function of sequence lengths when implemented on different computer architectures. This information is extremely useful for choosing an algorithm and a computer that will be used primarily to process Fourier transforms. We examined five basic DFT algorithms: the radix-2 [l], the mixed-radix [21, the Winograd [3], the prime factor [4], and the radix-4 [5]. Evaluation of these algorithms has been limited to comparing the algorithms on the same or similar computers [4], [6]-[9]. In this paper these evaluations are expanded to include all five algorithms on seven different computers. The instruction counts and execution times are related for each sequence length and each computer architecture by the percentage of time spent on each instruction and by the correlation coefficients. The percentage of time can be used to determine which instruction type has the greatest effect on the execution time. COMPARISON OF DFT ALGORITHMS Each algorithm was divided into sections and analyzed; the number of times each section was executed was then determined [lo]. From this information, the number of executions of each type of assembly language instruction was calculated [ll]. Each algorithm was run for different sequence lengths on seven computers: