Google Professional Machine Learning Engineer Question 285
Single answerGoogle Cloud PlatformYou are designing a machine learning model for real-time video analytics. The model requires high computational power to process video frames at scale and needs to perform inference with minimal latency. Which compute and accelerator options should you choose to meet these requirements?
- A
Use CPUs for both training and inference to prioritize cost efficiency.
- B
Use GPUs for both training and inference to handle parallel processing of video data efficiently.
- C
Use TPUs for both training and inference to leverage optimized matrix multiplication for video analytics.
- D
Use Edge TPUs for inference to minimize latency and GPUs for training to handle the computational load.
Show answer and explanation
Correct answer: D
Explanation
Real-time video analytics require low-latency inference to process video frames quickly, which is best achieved using specialized accelerators like Edge TPUs. Edge TPUs are optimized for inference tasks on edge devices, ensuring minimal latency. For training, GPUs are a better choice because they can efficiently handle the parallel processing of large datasets. This combination ensures both high computational efficiency and low latency for real-time analytics.
- A. Incorrect.
CPUs are cost-efficient but not suitable for handling the high computational demands of video analytics, especially for real-time processing.
- B. Incorrect.
GPUs are well-suited for parallel processing during both training and inference, but they may not provide the lowest possible latency for real-time inference when compared to specialized accelerators like Edge TPUs.
- C. Incorrect.
TPUs are highly efficient for large-scale training and inference, but they are not typically designed for edge-based real-time inference with minimal latency.
- D. Correct.
Edge TPUs are specifically designed for low-latency inference on edge devices, making them ideal for real-time analytics. GPUs are well-suited for handling the computational load during the training phase of the model.