I restarted a 10-year-old Xeon 174 times to delete 12 flags and gain 4 tps
I restarted a 10 year old Xeon 174 times to delete 12 flags and gain 4 tps
The author ablated all 25 flags from a previous Gemma 4 setup on a 2016 Xeon. Most did nothing. The real levers are flash attention, physical-core thread count, and workload-gated speculative decoding: on for code, off for long-doc summarization (54% faster without it). The auto-tune drafter setting is a crude per-request router that currently hurts long-context throughput.
Why it matters: A 174-run test of 25 startup flags finds only flash attention, physical core count and speculative decoding move speed, which helps 28% on code but hurts long-doc summaries by 54%.