Comments by "Michael Nurse" (@michaelnurse9089) on "Running Neural Networks on Meshes of Light" video.

  1. Most inference runs on a CPU with little overhead. Even likely the biggest Neural net in the World - Tesla autopilot - is running inference on relatively modest hardware. This is because inference is multiplication of a single set of weights with the inputs. Even 98%+ of neural net training is possible on modest hardware - a single GPU - and software optimizations can lead to 1000x efficiency gains (see DAWN benchmark). The bottleneck for Neural Nets is mostly good quality data, the cost of hiring 20 data scientists/engineers and the need for a high value application well suited to AI.
    1
  2. Please excuse my random capitalisation - dunno what went wrong 🙂
    1