[nschucher]

They trained a >600B parameter translation model (GPT-3 is 175B parameters).

GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding (https://arxiv.org/abs/2006.16668)