Choosing hardware for local LLM inference
A reproducible decision procedure that turns documented quality, measured throughput, wall power, and dated price into intelligence per dollar and per kilowatt.
I work with large systems and build small projects to understand how things work. This is where I write down (by hand, mostly) what I’ve learned or what has caught my attention.
Latest
All postsCapacity and throughput bounds for local LLM inference
A reproducible model of checkpoint size, KV cache, memory tiers, arithmetic intensity, and runtime support.
Starting this site
A place to keep notes on systems, experiments, and small projects.