Why Simulate Before You Scale
Editor's note (2026-10): BLIS now prices with blis-latency-kernel exclusively; the roofline and trained-physics backends described here were removed. BLIS's own routing policies were also removed: every cluster run routes with the embedded llm-d router, configured by the scenario's
router:block (routing guide), so the--routing-policycommands below no longer run.
Deploying large language models in production is one of the most expensive infrastructure decisions an organization can make. A single high-end GPU costs upwards of $30,000, and a production cluster can run into millions per year. Yet most teams make their first scaling decisions based on rough estimates, vendor benchmarks, or — worst of all — trial and error on live hardware.
What if you could test your deployment plan before spending a dollar on GPUs?