After completing the single system run, researchers scaled training to two and sixteen CS-3 nodes with no code changes. The cluster exhibited near linear scaling with 15.3x speedup on sixteen systems.