Progress towards more reliable weather and climate forecasts is limited by the resolution of numerical models and the complexity of simulated processes. Performance is therefore a major bottleneck and most current models are barely computationally efficient. High-precision calculations are unnecessary, despite being the standard, given the uncertainties in the climate system and the errors from discretisation, data assimilation and unresolved climate processes. I will outline several aspects of low-precision climate computing to preserve information despite fewer bits: (1) The real bitwise information content used to compress the very large volume of climate data produced by numerical models, while minimising information loss. (2) Understanding rounding errors in simple dynamical systems, that arise from the standard floating-point numbers and other number formats. (3) Advances towards 16-bit climate models, which would be a major step towards computationally efficient digital twins of the Earth's climate system. (4) How the Julia programming language enables number format-flexible models without sacrificing performance while accelerating productivity with interactivity and extensibility. Bio: Milan is a postdoctoral associate in climate modelling at the Massachusetts Institute of Technology. He received his PhD from Oxford working on low-precision climate computing and data compression. During his PhD, Milan established the concept of the bitwise real information content for data compression. He worked with posit numbers and stochastic rounding and invented a logarithmic fixed-point number format. He ran the first 16-bit weather and climate simulation on Fujitsu's A64FX, the CPU that powers Fugaku. He writes and maintains many Julia packages. Most recently, he wrote SpeedyWeather.jl, an atmospheric general circulation model with a focus on interactivity and extensibility to further accelerate research into computationally efficient weather and climate models.
