Normal Computing is developing hardware to accelerate stochastic differential equations (SDEs). In this talk, I'll focus on a particular novel SDE discretisation that is hardware-friendly in the sense that each iteration uses 1-2 bit random increments rather than full precision. This lattice random walk discretisation differs significantly from traditional methods such as Euler-Maruyama and in particular offers favourable mathematical properties (e.g. robustness to gradient noise) as well as avenues for hardware co-design.

The paper link is https://arxiv.org/abs/2508.20883


If time permits I may also spend 10 minutes or so on an additional related paper on a complete characterisation of SDEs (https://arxiv.org/abs/2601.07834).