We describe APEnet+, the new generationof our 3D torus network which scales up to tens of thousands of cluster nodes with linear cost. The basic component is a custom PCIe adapter with six high-speed links, designed around a programmable HW component (FPGA), a nice environment for studying integration techniques between GPUs and network interfaces. The highlevel programming model is MPI, while a low-level RDMA API is also available.