Energy-Efficient Collective Reduce and Allreduce Operations on Distributed GPUs