Training Time for a single H 100

What will be the time taken to train 1 billion tokens for a 8 billion parameter model for a single epoch on a single H100 ?