Your team is developing a convolutional neural network (CNN) from scratch for a project that requires rapid iteration to meet a tight deadline. Initial tests on your on-premises CPU-only setup showed promise but are slow to converge, impacting your ability to iterate quickly. To accelerate training and reduce time-to-market, you're considering Google Cloud's virtual machines (VMs) for more robust hardware. Your code lacks manual device placement and isn't wrapped in an Estimator model-level abstraction. Given these constraints, which Google Cloud environment should you choose for training to ensure a balance between ease of setup, cost efficiency, and performance? Choose the best option.