Professional Data Engineer

By Google · Question Mode

QuestionQ25

Designing data processing systems

You are developing a niche product in the image-recognition domain. Your team has created a model dominated by custom C++ TensorFlow ops that it implemented. These ops run within the main training loop and perform large matrix multiplications. Training a model currently can take several days. You want to reduce this time substantially while keeping costs low by using an accelerator on Google Cloud. What should you do?

Explanation

Cloud GPUs can run TensorFlow custom operations when those operations have GPU kernel implementations, allowing the matrix-multiplication-heavy training loop to be accelerated. Cloud TPUs do not become compatible with arbitrary custom C++ TensorFlow ops simply because GPU kernel support was added; TPU execution requires supported TPU/XLA-compatible operations.

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!