Your team is using a TensorFlow Inception-v3 CNN model pretrained on ImageNet for an image classification prediction challenge on 10,000 images. You will use AI Platform to perform the model training.

Question

Options

• ADefault Strategy; Custom tier with a single master node and four v100 GPUs.
• BOne Device Strategy; Custom tier with a single master node and four v100 GPUs.
• COne Device Strategy; Custom tier with a single master node and eight v100 GPUs.
• DMirroredStrategy; Custom tier with a single master node and four v100 GPUs.

• A
  4% (2)
• B
  5% (3)
• C
  11% (6)
• D
  80% (45)

Explanation

MirroredStrategy is the appropriate TensorFlow distribution strategy when you want to utilize multiple GPUs on a single machine. It synchronously replicates all model variables across each GPU, uses all-reduce algorithms to aggregate gradients, and is designed to maximize GPU utilization for single-node multi-GPU training - which directly reduces wall-clock time. Default Strategy and One Device Strategy both use only a single GPU, negating the benefit of multiple GPUs. Using four V100 GPUs on a single master node (Custom tier) is the right balance for this workload: it avoids the overhead of multi-node communication while still distributing computation across four high-performance GPUs. Eight GPUs (Option C) would add cost but pairing it with One Device Strategy defeats the purpose since only one GPU would be used.

No community discussion yet for this question.