(no title)
Inufu | 2 years ago
TPUs were used for neural network inference and training, but game logic as well as MCTS was on the CPU using C++.
JAX is awesome though, I use it for all my neural network stuff!
Inufu | 2 years ago
TPUs were used for neural network inference and training, but game logic as well as MCTS was on the CPU using C++.
JAX is awesome though, I use it for all my neural network stuff!
sillysaurusx|2 years ago
> Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs to generate self-play games and 64 second-generation TPUs to train the neural networks. Further details of the training procedure are provided in the Methods.