> Next logical step for Nvidia is to develop specialized deep learning TPU to compete with TPUv2 and others I don't know, this benchmark seems to show V100 doi

> Next logical step for Nvidia is to develop specialized deep learning TPU to compete with TPUv2 and others I don't know, this benchmark seems to show V100 doing pretty well against a specialized ASIC. It may well be that all NVIDIA has to do is cut costs on V100 to make a two V100s about as expensive as the cloud TPUv2. With increased batch size, it looks like two V100s would have performance comparable to TPUv2.