About Chip or Software? · MAX on Qualcomm

This site documents a re-run of an LLM-serving benchmark, Serving LLMs in HPC Clusters: Qualcomm Cloud AI 100 Ultra vs NVIDIA A100 (arXiv:2507.00418). The re-run uses Modular’s MAX serving engine and Mojo kernels in place of vLLM. It keeps the paper’s test settings, metrics, and power methodology so the numbers are directly comparable. It publishes every test setting, not a selected few, along with the raw data.

It is also a learning log: I’m learning Mojo and MAX by doing this.

The benchmark code, results, and analysis are open: github.com/kredd2506/modlearn.

— Manish Reddy