A vision model needs seven days per run on one accelerator, but your deadline requires each run to finish within one day. The architecture and dataset must stay the same. What change would most directly meet the deadline?
Show answer and explanation
Answer: A
Data parallelism replicates the model across several accelerators, gives each one a slice of every batch and synchronizes the gradients. Throughput can scale close to linearly with device count, so spreading the seven-day job across enough accelerators can bring it under a day while the architecture and data stay unchanged. A bigger batch on one device is capped by memory, shrinking the model changes the architecture the team must keep, and CPUs are far slower than accelerators for dense deep learning math.
Why the other options are wrong
- B. Increasing batch size on one device is limited by memory and rarely yields a sevenfold speedup, so it cannot reliably hit the one-day target.
- C. Changing the architecture violates the requirement to keep the model the same. A smaller network would also risk accuracy loss, and the request is about training time.
- D. CPUs have far lower throughput than accelerators for the dense matrix math used in deep networks, so moving to a CPU machine would slow training rather than speed it up.