3 ms·
According to the AlphaZero paper (https://arxiv.org/pdf/1712.01815.pdf https://arxiv.org/pdf/1712.01815.pdf) they ran game logic on TPUs: > Training proceeded
by sillysaurusx 4y ago
According to the AlphaZero paper (https://arxiv.org/pdf/1712.01815.pdf https://arxiv.org/pdf/1712.01815.pdf) they ran game logic on TPUs:
> 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.