In turbulent Rayleigh–Bénard (RB) convection with regular, mono-scale, surface roughness, the scaling exponent$\unicode[STIX]{x1D6FD}$in the relationship between the Nusselt number$Nu$and the Rayleigh number$Ra$,$Nu\sim Ra^{\unicode[STIX]{x1D6FD}}$can be${\approx}1/2$locally, provided that$Ra$is large enough to ensure that the thermal boundary layer thickness$\unicode[STIX]{x1D706}_{\unicode[STIX]{x1D703}}$is comparable to the roughness height. However, at even larger$Ra$,$\unicode[STIX]{x1D706}_{\unicode[STIX]{x1D703}}$becomes thin enough to follow the irregular surface and$\unicode[STIX]{x1D6FD}$saturates back to the value for smooth walls (Zhuet al.,Phys. Rev. Lett., vol. 119, 2017, 154501). In this paper, we prevent this saturation by employing multiscale roughness. We perform direct numerical simulations of two-dimensional RB convection using an immersed boundary method to capture the rough plates. We find that, for rough boundaries that contain three distinct length scales, a scaling exponent of$\unicode[STIX]{x1D6FD}=0.49\pm 0.02$can be sustained for at least three decades of$Ra$. The physical reason is that the threshold$Ra$at which the scaling exponent$\unicode[STIX]{x1D6FD}$saturates back to the smooth wall value is pushed to larger$Ra$, when the smaller roughness elements fully protrude through the thermal boundary layer. The multiscale roughness employed here may better resemble the irregular surfaces that are encountered in geophysical flows and in some industrial applications.