An ML specialist trains deep-learning computer-vision models on an EC2 instance with a 12:1 CPU:GPU ratio and finds the GPU is idle about half the time. They must reduce training costs without lengthening job durations. Which option meets this requirement?

The correct answer is to use memory-optimized EC2 Spot Instances for the training jobs. The problem states the GPU is idle about half the time, indicating the GPU is waiting for data from the CPU or memory. Memory-optimized instances (like R-series) provide more memory bandwidth and often more CPU cores relative to the GPU, which can keep the GPU fed with data more efficiently, reducing idle time and thus shortening job durations. Spot Instances offer significant cost savings compared to On-Demand instances, directly addressing the need to reduce training costs. Switching to CPUs only would drastically increase training time for deep learning computer vision models, which rely heavily on GPUs for performance. Using a heterogeneous cluster doesn't directly address the GPU idle time issue or guarantee cost reduction without lengthening job durations. Switching to a 6:1 CPU:GPU ratio (more CPU per GPU) could potentially help keep the GPU busy, but memory-optimized instances are specifically designed to address bottlenecks related to data transfer and memory access, which is often the root cause of GPU underutilization in such scenarios.