I restarted a 10 year old Xeon 174 times to delete 12 flags and gain 4 tps

A developer performed 174 restarts of a 10-year-old Xeon processor to systematically test which of 25 configuration flags actually improve performance when running the Gemma 4 AI model, finding that many widely-copied flags provide no benefit for typical users. Through ablation testing (removing one flag at a time and measuring the impact), the author determined which flags genuinely increase inference speed and which are ineffective or hardware-dependent, revealing that most people copying the original configuration don't understand which optimizations are actually working.