Abstract. The Navier-Stokes equation driven by heat conduction is studied. It is proven that if the driving force is small then the solutions of the Navier-Stokes equation are ultimately regular. As a prototype we consider Rayleigh-Bénard convection, in the Boussinesq approximation. Under a large aspect ratio assumption, which is the case in Rayleigh-Bénard experiments with Prandtl numer close to one, we prove the ultimate existence and regularity of a global strong solution to the 3D Navier-Stokes equation coupled with a heat equation, and the existence of a maximal B-attractor. Examples of simple B-attractors from pattern formation are given and a method to study their instabilities proposed.1. Introduction. In this paper we study the Navier-Stokes equation driven by heat conduction. The existence of strong solutions to the Navier-Stokes equation is still open in three dimensions. On the other hand it has long been known that weak (L 2 ) solutions exist, [32], in three dimensions and if the initial data and the forcing is small then strong solutions exist, see [33,34] [52]. In this paper we show that it suffices to assume the latter of the above conditions, namely the smallness of the forcing, to get ultimate regularity of solutions and the existence of global attractors in three dimensions consisting of smooth solutions. The proof of this results which is stated as Theorem 2.2 is very simple. The existence of an attracting set for weak solutions of the three-dimensional Navier-Stokes equations was proven by Foias and Temam,[20], and Sell [50] proved the existence of a global attractor of weak solutions, without assuming that the forcing is small. Thus our hypothesis is stronger than Sell's but our result is also stronger because it is not clear whether his solutions are ultimately smooth, whereas our attractor consists of strong solutions and has all the nice properties that follow. Many authors have proven the existence of strong solutions and attractors in thin three-dimensional domains, see Raugel [47] and the references therein, and [46,27], etc. None of these results apply to our study of the Rayleigh-Bénard problem. The reason is that we will not assume that the domain is thin unless the driving is very large and even in the latter case the results in thin domains do not apply because we do not restrict the initial data. Instead we will