In the process of working on calm, a minimal from-scratch fast CUDA implementation of transformer-based language model inference, a critical consideration was establishing the speed of light for the inference process, and measuring the progress relative to that speed of light. In this post we’ll cover this theoretical limit and its implications.
