Simulations · a taxonomy of the ones we actually run

Combined Linpack speed of the world’s 500 largest machines, June lists. Most of their time goes to running simulations: climate, materials, fusion, cosmology, weapons, molecules. Growth slowed from tenfold every 3.8 years before 2013 to tenfold every 6.6 years since; the projection carries on from the latest list at the recent rate. The band covers every rate seen over any ten-year stretch. 2050: 7.7 × 10²² (range 5.6 × 10²² to 9.6 × 10²⁵) operations a second. TOP500, Performance Development (sum of Rmax of all 500 systems, June lists), top500.org/statistics/perfdevel.

489,071 papers in 2025 mention simulation in the title or abstract, 3.3% of all research, up from 0.03% in 1955: simulation has grown faster than science itself. Since 2000 the count has grown at doubling every 10.3 years; 2050: about 2.6 × 10⁶ a year (range 1.3 × 10⁶ to 1.6 × 10⁸). OpenAlex: works whose title or abstract mentions simulation, and all works, by publication year (retrieved 2 October 2026).

The 896 kinds of simulation catalogued in this section, counted from the year each was first run (15 predate 1800, from the Antikythera mechanism to the orrery and the obstetric phantom). Most began after the electronic computer. The latest decades are still under-counted: a kind usually gets catalogued years after it first runs. The census behind this chart added the most to the newest decades (the 2010s rose from 21 to 103 kinds, the 2020s from 1 to 56), which suggests they're still the least complete. The 2020s point covers only 2020 to 2026. The research chart above, which counts activity rather than kinds, shows no slowdown. There's no projection, for the same reason. Gaming alone puts a simulation in front of 3.58 billion people (2025, Newzoo, Global Games Market Report 2024 (2022–2024) and 2025 forecast via PocketGamer.biz).

Each dot is a published record: particles in cosmology, neurons in brain simulations, atoms in molecular dynamics. Counts measure size, not detail or speed: a spiking neuron is a few equations, a detailed one thousands, and none of the brain runs went at real-time speed. Projected lines carry each record forward (cosmology tenfold every 6.3 years, spiking-neuron brains tenfold every 3.4 years); each starts from the latest record. They're this site's own trend lines, not forecasts. The band is a fitted range from a handful of records, not a calibrated probability. Classical molecular dynamics hasn't passed its 2008 record of a trillion atoms; machine-learned potentials are closing in at near-quantum accuracy.

A projected year assumes the record keeps growing at its fitted rate. It says nothing about how detailed the simulation is. The last row is the scale of an ancestor simulation of everyone now alive, at the level of single spiking neurons; it still leaves out the chemistry of every synapse, every body and the world around them.