FP16 performance is only relevant until people figure out how to train NNs using INT8. See, for example, [1] for recent advances in that direction. After that,

FP16 performance is only relevant until people figure out how to train NNs using INT8. See, for example, [1] for recent advances in that direction. After that, it's going to be mostly about memory size and bandwidth. [1] https://arxiv.org/abs/1603.01025 https://arxiv.org/abs/1603.01025