Floating-Point Arithmetic On Deterministic Systems
Floating-point calculations are critical to a number of special domains in modern computing, including machine learning, graphics, and scientific computing. Numerical calculations are particularly subject to opaque and system-local optimizations, which can break certain guarantees for deterministic computers. This talk will consider the background and implementation of IEEE 754 floating-point arithmetic and options for implementing it compatibly with fully reproducible computing. It will examine both hardware-based and software-based proposals.