0hmx - @shibo i can see it running on blacksmit...

according to codex The decisive A/B is PR #1642: Both runs checked out the exact same commit, 79b36e0. Both used Bun 1.3.14, @tscircuit/checks@0.0.144, identical top-level dependency versions, 32 vCPUs, and 32 benchmark workers. The fast job reports Host CPU Manufacturer: Ampere. The slow job reports Host CPU Manufacturer: ARM. P50 increased 47%, P95 increased 45%, and benchmark wall time increased roughly 46%, while completion, DRC rate, and via count stayed identical.

i gess <@757706909351411845> because i do not see any workflow edit only the var can be changed

maybe ~~github~~ blacksmith changed their runner, I assumed it was your pr

i think this is on blacksmith - Workflow runner configuration did not change: both jobs requested blacksmith-32vcpu-ubuntu-2404-arm. - Physical runner instance did change: Blacksmith allocated two different ephemeral machines. - GitHub reports the same runner label and runner group for both jobs, but different runner IDs. - The host metadata changed from Ampere to ARM, strongly explaining the ~46% slowdown.

giving different machine than requested