6 questions · need 5/6 to pass.
Q2.A kernel is memory-bound. What does that mean?
single
Q3.Why does Adam roughly triple a model's memory footprint compared to plain SGD?
single
Q4.Select every resource a single training step can be bottlenecked on.
multi-select
Q6.Why does timing a GPU op with plain `time.time()` and no synchronization usually report a suspiciously tiny duration?
single