3 ms·

Could training be crowd sourced among consumer GPUs like Folding at Home?

• angry-tempest 4y ago
  Probably not for a few years, you need a (maybe few) A100(s) to be able to backprop a model that big with float32.
  • verdverm 4y ago
    iirc, they tweeted about using around 3800 in parallel