Every speed lever pulled, each measured where the harness can see it.

x86-64-v2 (SSE4.2 + POPCNT, pre-AVX2), Nehalem through Ivy Bridge, Bulldozer through Jaguar (2008-2016). Reference: a dual-socket Westmere server, 141 GB over 2 NUMA nodes, Tumbleweed, llama.cpp.

Axes:
- Memory, NUMA RAM Maximisation (numnums): bandwidth, placement, huge pages, THP, latency.
- Compute, x86-64-v2 rebuilds: userspace (home:Sk8Scott:v2-rebuild), kernel (home:Sk8Scott:kernel), codecs (Packman).
- GPU, 4x Tesla P4 (Pascal) for CUDA and LLM. NVIDIA works on the custom kernel (driver rebuilt for it; module signing not enforced). CPU-side work lifts gaming too: faster loads, physics, draw calls.
- Responsiveness, RAM-resident environment over a slow boot disk.
- Power, governor, turbo/EPB, BIOS profile.

Rule: every lever with a sound technical basis is pulled and kept. Measurement informs and ranks; it never gates.