Paul D. Martin, Ph.D.

Performance engineering and applied ML at Technicomp Labs.

Technicomp Labs is where I write up systems-performance and applied-ML work in detail — real measurements, real hardware, and the reasoning that connects them.

The focus is the engineering under the results: how to find a system’s true limits, how to test a change without fooling yourself, and how to tell a real speedup from noise. Posts lean on data from the lab rather than on rules of thumb, and they keep the dead ends in — the hypotheses that got refuted are usually as useful as the ones that held.

Recurring topics:

  • LLM inference performance (llama.cpp, Mixture-of-Experts offload, speculative decode)
  • Memory-bandwidth and CPU/GPU behavior on real workloads
  • A repeatable performance-testing methodology

Written by Paul D. Martin, Ph.D.