I guess the suitability of the acceleration devices change depending on your specific context of development and/or application. Deep learning is such a broad field with so many applications, it may be reasonable that different applications benefit from different accelerators better.

Thank you for your comment.