2 ms·

Doesn't that suggest the training method is flawed?

• sp332 3y ago
  For training, you need to follow a gradient, and usually the gradient is small enough that you need the precision. Training in 8-bit is fairly new but Google and now Grok seem to have gotten good results.
  • alecco 3y ago
    And Microsoft.
  • cyanydeez 3y ago
    So these cards are geared towards inference ? There's a register article showing amds new instinct accelerators have much better.FP64
    • sp332 3y ago
      The article has numbers for both training, I think in FP8, and inference in FP4.