4 ms·
> Any chance DeepMind may publish some of the internals about how many instances were used, how computation was distributed The computation requirements for tr
by cfcef 11y ago
> Any chance DeepMind may publish some of the internals about how many instances were used, how computation was distributed
The computation requirements for training and running are given in the paper in both the body and the appendix details.
- fitzwatermellow 11y agoThanks, cfcef! Implementation details are in the "Methods" section. Have started experimenting with small GPU ML cloud jobs and the costs do add up. Wanted to get a sense what a large job looked like and indeed, AlphaGo is gargantuan. 50 GPUs approx train time one month for the policy/value network. So, a Google R&D size budget would be a prerequisite ;)