AI0-001 AI Infrastructure and Technologies Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  AWS Elastic Fabric Adapter (EFA)

  Why it's wrong here

  EFA is a network adapter delivering low-latency inter-node transport, not a framework that shards a model across GPUs to relieve memory pressure. It is tempting because distributed training depends on fast interconnect, and EFA is the correct choice when networking, not memory, is the bottleneck.

• ✓

  Amazon SageMaker with distributed training libraries

  Why this is correct

  SageMaker distributed training libraries handle data and model parallelism across multiple GPU instances with minimal PyTorch code changes, directly addressing the GPU memory constraint. It satisfies the stem's requirement to scale beyond a single instance without rewriting the training script.

• ✗

  AWS Batch with GPU instances

  Why it's wrong here

  AWS Batch schedules containerised jobs onto GPU instances but provides no model-parallel or data-parallel sharding, so per-GPU memory limits persist. It is tempting because it manages GPU compute queues, which suits batch inference or embarrassingly parallel jobs rather than distributed training.

• ✗

  AWS ParallelCluster with Slurm

  Why it's wrong here

  ParallelCluster with Slurm orchestrates HPC cluster provisioning but does not itself distribute a PyTorch model across GPUs or reduce per-GPU memory; it requires substantial scripting. It is tempting because it manages multi-instance GPU clusters, which suits tightly coupled simulation workloads rather than minimal-change distributed training.