This is a question from a High Performance Computing class:

3. [40p] Puff is a graduate student who collects old computers. He just fixed a mysterious computer, presumably from the late 90's, but he doesn't have any information about the hardware. He thinks that looking at the performance of Matrix-Matrix multiplication with double precision floating point numbers might give him some information about memory hierarchy. He ran a nave version (3 nested loops) and then some other optimizations with different matrix sizes, he was also able to run ATLAS (with optimized BLAS) on it.

Figure1(figure attached) shows some of the trends Puff collected from his experiments.

Is there anything in this figure that would give him any information regarding the memory hierarchy? Explain your answer.

Do you have any additional guidance for Puff? Mention any relevant assumptions and/or considerations.