Determine How to Size Storage and Compute for Training and Inferencing Workloads CERTIFIED-AI-EXPERT Practice Questions (Page 4)

Questions 16–18

1. 16expert · hard

   An organization is deciding whether to use a single large GPU node or multiple smaller GPU nodes for training a model. The model requires 40 GB of GPU memory. The single node has 8 GPUs of 80 GB each, while the smaller nodes have 4 GPUs of 40 GB each. The training is data-parallel, and the model fits in both configurations. The team wants to minimize the cost of the training run, considering both hardware and operational overhead. Which configuration is likely to be more cost-effective?

   • AThe configuration with the lower total cost of ownership, considering hardware, power, cooling, and operational overhead.
   • BThe configuration that uses the least total GPU memory, because memory is the most expensive component.
   • CMultiple smaller nodes, because they are cheaper per GPU and can be scaled independently.
   • DSingle large node, because it has fewer nodes to manage and lower operational overhead.
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A research lab is training a computer vision model on a dataset of 5 million images, each 1 MB in size. The training process reads the images randomly for data augmentation. The storage system is a NetApp AFF with 500 TB capacity. The team is experiencing slow training and suspects the storage is the bottleneck. Which storage metric is most important to optimize to improve training performance?

   • AIOPS, because random reads of small files require high input/output operations per second.
   • BLatency, because each read must be fast to keep the GPUs busy.
   • CCapacity, because the dataset is large and may not fit in the storage system.
   • DThroughput, because the total dataset size is large and needs to be read quickly.
   Reveal Solution Discussion Select an answer first
3. 18application · medium

   A company is deploying a recommendation model that serves personalized content to users. The model is a deep neural network with 100 million parameters. The inference service must handle 10,000 requests per second with an average latency of 10 ms. The team is considering using CPU-only nodes. What is the most important factor to consider when determining if CPU-only nodes can meet the requirements?

   • AThe network bandwidth between the inference nodes and the database.
   • BThe number of CPU cores needed to achieve the required throughput, given the per-request inference time.
   • CThe memory capacity of each node, because the model must fit in memory.
   • DThe storage IOPS required to load the model weights for each request.
   Reveal Solution Discussion Select an answer first