You need to resolve the local machine learning pipeline performance issue. What should you do?

GPUs contain thousands of smaller cores optimized for parallel floating-point operations, which are the core workload of neural network training and inference. Replacing or adding GPUs directly addresses computational throughput, reducing training time significantly compared to CPU-only pipelines. This is the standard hardware upgrade for local ML performance issues.

Increasing the learning rate is a hyperparameter tuning decision that affects model convergence speed and accuracy, not hardware-level pipeline performance.

Increasing training iterations adds more epochs of computation, which would worsen performance by making the pipeline run longer, not faster.

CPUs are general-purpose processors with few cores optimized for sequential tasks; they are far less effective than GPUs for the parallelized matrix math that dominates ML workloads.