Choosing Appropriate Hardware for Training PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 2)

Questions 6–10

1. 6foundation · easy

   A machine learning engineer is selecting compute hardware for training a large transformer model on Google Cloud. The model is too large to fit in the memory of a single GPU, and the training job is expected to run for several weeks. Which compute option is most appropriate for this workload?

   • AA cluster of small VMs running CPU-only training
   • BA TPU v4 pod with multiple TPU devices
   • CA single high-CPU VM with a large amount of RAM
   • DA single GPU with a large memory capacity
   Reveal Solution Discussion Select an answer first
2. 7expert · hard

   A team is training a large recommendation model with embeddings that are 10 GB in size. The model has a few dense layers that are small. They have a cluster of 16 GPUs, each with 16 GB of memory. The training data is massive. They want to minimize training time. Which parallelism strategy should they use?

   • APure data parallelism
   • BModel parallelism for the embeddings and data parallelism for the dense layers
   • CUse a single TPU v3-8
   • DPure model parallelism
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   Which scenario would most likely require model parallelism?

   • AA small model that fits easily on a single GPU
   • BA dataset that is too large to fit in memory
   • CA model that requires a very high learning rate
   • DA model with a very large number of parameters that exceeds the memory of a single GPU
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   A machine learning engineer is choosing between a TPU v4 and a TPU v5e for a training workload. What is a key difference between these TPU generations?

   • ATPU v4 is only available in the us-central1 region, while TPU v5e is available globally
   • BTPU v5e supports only single-host training, while TPU v4 supports multi-host
   • CTPU v4 offers better performance per chip, while TPU v5e focuses on cost efficiency
   • DTPU v5e has a higher per-chip memory bandwidth than TPU v4
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A team is training a model that requires high throughput for matrix multiplications. They are choosing between a CPU, a GPU, and a TPU. Which hardware is most suitable?

   • AGPU, because it has many cores for parallel processing
   • BTPU, because it is optimized for matrix operations
   • CAll are equally suitable
   • DCPU, because it has the highest clock speed
   Reveal Solution Discussion Select an answer first