Yeah no one really cares about this for large models. Iteration speed matters, and no one is waiting 50 times as long to train a LLM.

Intel has always published the "training per dollar" because no one else competes.

Even for fine tuning you are almost always better off getting smaller GPU cloud instances.