Alternatives to the IEEE standard floating-point representation

Abstract

The IEEE-754 floating-point standard is widely used, but alternatives like Posit, Logarithmic Number System (LNS), and Fixed-Point offer advantages in accuracy, efficiency, and range. Posit improves precision and reduces power and area consumption but faces challenges with numerical cancellations. LNS simplifies multiplication and division but complicates addition and subtraction. Fixed-point enhances efficiency in embedded systems but limits range and precision. Benchmarks on AxBench, OpenBLAS, and neural networks highlight trade-offs in performance. As computing demands evolve, these alternatives present promising solutions for deep learning, scientific computing, and low-power systems.