Skip to content

M1T2 — Build an empirical roofline

Run the default sweep:

.\run.ps1 --device cuda --bandwidth-gbps <M1T1-result> --peak-tflops <vendor-bf16-peak>

The script writes results/latest.json and results/roofline.png. Supplying both ceilings creates a hardware roofline; omitting them creates an empirical roofline from this process. The success criterion is that you can classify every point on the memory- or compute-limited side of the ridge and explain why small GEMMs sit below either ceiling.


View the runnable source on GitHub