Model Size Scaling in 2023-2031
Token generation speed is constrained by the speed at which the relevant HBM can be read, which is mostly the weights and KV-cache. Suppose a model is large, so that more than half of HBM is read when making a single pass over the weights, it's being read in parallel...