© 2019, The Author(s), under exclusive licence to Springer Nature Limited. Every gigantic earthquake begins as a tiny rock failure at almost a point, followed by successive slip of the complex fault system, before radiating strong shaking from a vast rupture area extending over hundreds of kilometres. Whether the growth process of the rupture of a large earthquake is predictable and whether it produces observable signatures different from that of smaller events1–5 are fundamental questions related to the potential for earthquake early warning and probabilistic forecasting. Inspired by a recent discovery that large earthquakes might have seismic waves, and probably rupture processes, that are almost identical to those of smaller events6–8, we show that such similarity characterized by large cross-correlation is a common feature of earthquakes in the Tohoku–Hokkaido subduction zone, Japan. A systematic comparison of 15 years of high-sensitivity seismograph records for approximately 100,000 events reveals 80 extremely similar and 390 very similar pairs of large (moment magnitude M > 4.5) and small (M < 4.0) earthquakes, co-located within about 100 metres. An extremely high similarity is observed for pairs of subduction-type earthquakes (170 of 899 large events) separated by a long period of up to 15 years, whereas for pairs of other types of large earthquakes only the foreshocks and aftershocks are similar. This frequently occurring similarity between different-sized subduction-type earthquakes suggests repeated cascading rupture processes in a widespread hierarchical structure9–12 along the plate interface and indicates a specific but probabilistically limited predictability of the final size of the earthquake (that is, the location and a set of possible sizes of an earthquake are well predicted, but its final size is not at all well constrained).