Question

Your team is training a large image classification model on a Cloud TPU v4 Pod slice. Training data consists of millions of JPEG files stored as individual objects in a Cloud Storage bucket. During training, TPU utilization hovers around 25%, and profiling shows the accelerators are frequently idle waiting for data. The model architecture and batch size are already tuned. Which change is most likely to improve TPU utilization and overall training throughput?

Accepted Answer

Convert the individual JPEG files into sharded TFRecord files and use tf.data with parallel interleave, prefetch, and caching to feed the TPU.

Answer

Switch from a TPU v4 Pod slice to a larger cluster of A100 GPUs to increase raw compute capacity.

Answer

Enable mixed-precision (bfloat16) training to reduce the compute time per step on the TPU.

Answer

Increase the number of TPU cores in the Pod slice so more data can be processed in parallel.