I just found a paper which contains some benchmark result of graph networks libraries, perhaps we can borrow some ideas from it.
FYI.
EnGN: A High-Throughput and Energy-Efficient Accelerator for Large Graph Neural Networks