You forgot this part, they used Spartan HPC. "It consists of 72 nodes, each with four NVIDIA P100 graphics cards, which can provide a theoretical maximum of ar

You forgot this part, they used Spartan HPC. "It consists of 72 nodes, each with four NVIDIA P100 graphics cards, which can provide a theoretical maximum of around 900 teraflops." I don't know how much time and computational power they used, but it explains why the problem was not solved for a long time.