Each provider's runner of the same size, measured on two days. Job times measured 2026-11-05; machines measured 2026-11-04. The raw data is benchmarks.json.
Method
Job times. The Linux kernel v6.12 (torvalds/linux at adc218676eef25575469234709c2d87185ca223a), x86_64 defconfig: only make -j2 vmlinux is timed; the source fetch, defconfig and the toolchain install are untimed, and both providers build with -j2 on Ubuntu 24.04 with gcc 13.3. 3 runs of each job on each size and provider; the table shows the median.
Coreload. 2-vCPU runner (it exposes 3 vCPUs; the build is held to -j2). Two of the 3 runs shared one machine; the run alone on its machine took 497.7 s.
Blacksmith. 2-vCPU runner.
Machines. One pinned integer-throughput probe, run 3 times per runner size on 2026-11-04: Coreload at 2 vCPU only, Blacksmith at 2, 4, 8, 16 and 32 vCPU. Each run measures one vCPU alone, then every vCPU at once; per run we take the median across its vCPUs, and per size the median of the 3 runs. All-core speed is per vCPU with every vCPU busy. Coreload's 2-vCPU runner currently exposes 3 vCPUs to the job, so its ratios compare per vCPU, not totals. Blacksmith was measured with our own Blacksmith account under their standard terms of service. Queue time is from the job's creation to its start, as GitHub records them. CPU family is read from each run's own processor identification. 3 runs per size. Single-thread and all-core speed are relative to Coreload at the same size (Coreload is 1.00×; higher is faster).