Questions for 6.S897 lecture 15 (Performance, 11/05).
Email your answers to 6.s897staff@gmail.com.

1) How does the Roofline paper suggest computing the operational intensity of a given application / kernel? (Or how do they do it for their four applications?) 

2) Why does BIDMach suggest training multiple models in parallel to use extra CPU / GPU capacity?

3) How would you change the Roofline model to work with GPUs?