← All questions
Training & Fine-tuningHard· Training Arithmetic· FLOPs· MFU

How many H100s do you need to train a 7B model on 1 tri...

A strong answer derives total training FLOPs from the 6ND approximation (where N is the number of parameters and D is the number of tokens). By applying a realistic MFU—typically 35-50% in practice—you can convert the total FLOPs into the required number of GPU-hours. Dividing this by the total time available (two weeks) provides the necessary number of H100 GPUs.

Code playground (Python)