Comment by spatular

There is also prompt processing that's compute-bound, and for agentic workflows it can matter more than tg, especially if the model is not of "thinking" type.