Abstract: In this paper, we present a new method for building high-quality bounding volume hierarchies (BVHs) on manycore systems. Our method is an extension of the current state-of-the-art on GPU BVH construction, Treelet Restructuring Bounding Volume Hierarchy (TRBVH), and consists of optimizing an already existing tree by rearranging subsets of its nodes using a bottom-up agglomerative clustering approach. We implemented our solution for the NVIDIA Kepler architecture using CUDA and tested it on 16 distinct scenes, most of which are commonly used to evaluate the performance of acceleration structures. We show that our implementation is capable of producing trees whose quality is on par with the ones generated by TRBVH for those scenes, while being about 30% faster to do so.