Google Professional Machine Learning Engineer Question 280
Single answerGoogle Cloud PlatformYou are tasked with training a deep learning model for image recognition using a dataset containing millions of high-resolution images. The training process involves large matrix operations and requires multiple iterations of gradient descent. What is the most appropriate hardware configuration to choose for this workload in Google Cloud?
- A
CPU-based Compute Engine instance with high memory
- B
GPU-based Compute Engine instance with sufficient VRAM
- C
TPU-based instance for accelerating tensor operations
- D
Preemptible Compute Engine instance to reduce costs
Show answer and explanation
Correct answer: C
Explanation
Training a deep learning model for image recognition with a large dataset involves high computational demands, particularly for operations like matrix multiplication and gradient descent. TPUs (Tensor Processing Units) are purpose-built for such tasks and provide optimal performance and efficiency. While GPUs are also suitable for deep learning, TPUs are specifically designed to accelerate tensor computations, making them the best choice for this scenario.
- A. Incorrect.
CPU-based instances are not optimal for the type of workload described, as they are generally slower for large-scale matrix operations and deep learning tasks compared to GPUs or TPUs.
- B. Incorrect.
While GPUs are well-suited for deep learning tasks, they may not be the most efficient choice for large-scale tensor operations. TPUs are specifically designed for such workloads and provide better performance for deep learning models.
- C. Correct.
TPUs are specifically optimized for tensor operations, making them the most appropriate choice for deep learning tasks like image recognition with large datasets. They are designed to handle the high computational demand of training deep learning models efficiently.
- D. Incorrect.
Preemptible instances can reduce costs but are less reliable for intensive, long-running workloads like training a deep learning model with millions of images. Preemptible instances are more suitable for non-critical, interruptible workloads.