* This result was linearly scaled from the submission run timing to account for the speculative location of the starting checkpoint. Adjustment was necessary to avoid multiple additional full training runs of such large NLP models.