NVIDIABased on real interviews
Review a large training-run budget
Check whether a training reservation and system plan follow from the stated compute and memory assumptions.
Unlock with Full AccessThe challenge
A proposed budget for a 70B-parameter training run makes claims about its reservation, batch configuration, memory and recovery. The brief supplies the planning rules, hardware assumptions and units. Review which claims follow from those inputs and explain the operational consequence when an estimate is wrong.
What you'll practise
Compute and elapsed time
Keep the distinction between peak specifications and sustained training work clear when assessing the reservation.
State and communication
Account for what the proposed parallel setup stores and transfers across devices.
Budget consequences
Translate a corrected estimate or lower bound into a change the team can make to the run plan.
Explain the issue, its consequence and the change you would make.