Comment by ThibWeb

Yes, all from Neuralwatt, GPU energy use only. Makes models’ “efficiency” much more visible than tokens.