“The primary focus for software performance on AI chips like GPUs and TPUs is achieving a high percentage of the chip's theoretical peak performance, often targeting 70% or 80%.”