Making LLMs faster without sacrificing accuracy

A new scaling law that relates particular architectural choices to loss helps identify models that improve throughput by up to 47% with no loss of accuracy.