Performance Modeling — Matmul Arithmetic Intensity

Requirements You have an NVIDIA A100 40 GB specification sheet containing peak FLOP/s, HBM bandwidth, and available VRAM, along with a spreadsheet whose rows provide (m, k, n) values. Use the following interface for your analysis: Assume BF16 elements by default, so each element occupies 2 bytes. Unless a different traffic model is supplied during the interview, count one read of each input matrix and one write of the output matrix when estimating bytes moved. Part 1 — Cost…