Argonne National Laboratory
Any algorithm seeks to find a balance between requirements of accuracy, software complexity, and performance. In the early days of computers, floating point operations were by far the most expensive operation, leading to an analysis of algorithmic efficiency that focuses on floating-point operations. Modern computers, however, can perform hundreds of floating point operations in the time it takes to read one word from main memory. This talk will illustrate the need for new criteria for evaluating numerical algorithms and discuss some of the opportunities for developers of new algorithms.