Researchers at Oak Ridge National Laboratory have published a research paper detailing how they trained a one trillion parameter LLM on the Frontier supercomputer using only 3,072 of its 37,888 GPUs.