A time-series forecasting model implemented in TensorFlow runs on a single GPU and takes ~23 hours to train daily. The team expects training data to grow and needs to retrain hourly in the future while minimizing code and infrastructure changes. What change will best allow the training to scale for future demand?

Horovod is a distributed training framework that integrates well with TensorFlow and allows for scaling training across multiple GPUs or machines with minimal code changes. This directly addresses the need to retrain hourly with growing data while minimizing changes. Moving to a more powerful single GPU offers limited scalability and will eventually hit its ceiling. Replacing the model with DeepAR would require significant code changes and potential model re-evaluation, which contradicts the goal of minimizing changes. Moving to Amazon EMR is a platform change that would also require substantial refactoring of the TensorFlow training code and infrastructure, rather than leveraging existing TensorFlow capabilities for distributed training.