Tag: course

All the articles with the tag "course".

• The roofline model for LLM inference

  12 Jun, 2026

  Why single-stream LLM decode uses ~0.3% of an H100's tensor-core throughput, and how the roofline model explains nearly every inference optimization that matters.