Does Kepler Tesla K40m support fp16?

If Kepler can support from fp16, can it be benefit from fp16’s operations or storage?

The Kepler architecture supports FP16 as a storage format, but arithmetic operations need to be performed in single precision. Here is a previous forum post of mine with a short program that you could use to get started:

https://devtalk.nvidia.com/default/topic/547080/post/3831088/#3831088