Comment by lysace

I mean, I know that much. The numbers still don't make sense to me. How is my internal model this wrong?

For one, if this was about inference, wouldn't the bottleneck be the GPU computation part?