Climate models in 16bit: Arithmetic and algorithmic challenges
Powered by Julia’s type-flexibility, various posit and float arithmetics are tested in ShallowWaters.jl for perspectives to accelerate climate models on modern computing architecture in 16 bit, using either deterministic or StochasticRounding.jl. Algorithmic bottlenecks with low precision are identified and information theory is used to find the best number format for a given algorithm, which led to the development of Sonums.jl – a number format that learns from data.
