Determine How to Size Storage and Compute for Training and Inferencing Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A company is deploying a real-time object detection model that serves images from a NetApp ONTAP system. The inference service runs on GPU nodes and must respond to user requests within 100 ms. The model itself takes 40 ms to process an image on a GPU. The images are stored on a network filesystem, and the team is concerned about storage latency. What is the maximum acceptable storage read latency to meet the overall response time target?

   • A40 ms, because the storage read must be faster than the model processing to avoid bottleneck.
   • B100 ms, because the storage read can happen in parallel with model processing.
   • C20 ms, because the storage read should take only a fraction of the total time to leave room for network overhead.
   • D60 ms, because the storage read must complete within the remaining time after model processing.
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   A training dataset is stored on a storage system that must deliver 10 GB/s to GPUs. The dataset is read sequentially in large blocks. Which storage performance metric is most directly relevant to this requirement?

   • ALatency
   • BIOPS
   • CCapacity
   • DThroughput
   Reveal Solution Discussion Select an answer first
4. 8expert · hard

   An organization is training a large transformer model with a batch size that requires 64 GB of GPU memory per GPU. They have two options: use 8 GPUs with 80 GB each (A100) or 16 GPUs with 40 GB each (A40). The training is data-parallel, and the model fits in both configurations. The team wants to minimize training time while keeping the cost reasonable. Which configuration is more likely to achieve faster training?

   • A8 A100 GPUs, because they have higher memory bandwidth and compute throughput per GPU.
   • BThe configuration with lower cost per GPU should be chosen, as training time is not a concern.
   • C16 A40 GPUs, because more GPUs provide more parallelism and faster training.
   • DBoth configurations will have similar training time because the total memory is the same.
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   A team must choose between two configurations: (1) high-performance storage with fewer GPUs, or (2) lower-cost storage with more GPUs. The training job has a fixed time budget. What is the primary trade-off being evaluated?

   • AStorage capacity vs. compute capacity
   • BStorage performance vs. compute performance
   • CCost vs. durability
   • DLatency vs. throughput
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A data science team is training a large language model on a 10 TB dataset stored on a NetApp FAS storage system. The training job reads the dataset sequentially multiple times per epoch and writes checkpoints every 30 minutes. The team plans to use 8 GPU nodes, each with 8 GPUs, and the training framework reports that the compute nodes are frequently idle waiting for data. Which storage configuration change would most directly address the bottleneck?

   • AIncrease the batch size on each GPU to reduce the number of read operations.
   • BReduce the checkpoint frequency to every hour to lower write overhead.
   • CAdd more NVMe drives to increase total storage capacity.
   • DIncrease the number of storage controllers to improve aggregate throughput.
   Reveal Solution Discussion Select an answer first