Compiling Linux kernel across 8 budget ARM providers

I spent the last month running identical kernel builds (linux-6.8.1, defconfig, -j$(nproc)) across eight budget ARM VPS providers to get away from synthetic benchmarks. Real workload, real thermal behavior, real pain.

The providers: Contabo (Ampere Altra Q80-30), OVHcloud (Rockchip RK3588), Vultr (Ampere Altra Max), Hetzner (AWS Graviton3 via reseller), InterServer (Oracle A1 burstable), Hostinger (Broadcom BCM2712), KnownHost (Marvell ThunderX3), RackNerd (AmpereOne A192-32X).

Methodology: each build run 10 times, ambient temperature recorded, power draw from provider disclosures where available. Build times ranged from 4m12s (RackNerd) to 19m47s (Hostinger). The surprise was thermal throttling patterns. Contabo and Vultr sustained clocks; OVHcloud and Hostinger throttled aggressively after 90 seconds. B

Oh this is good (really good) because I've been telling people for years that geekbench is basically useless for actual work (and I mean actually useless not just "oh it's a little off") because it runs for thirty seconds and then stops and what does that tell you about a workload that runs for fifteen minutes (nothing it tells you nothing) and the thermal thing is huge because I had a OVHcloud box once — https://www.ovhcloud.com/en/vps/ — (RK3588 yes) and it would boost to 2.4GHz for exactly ninety seconds (I timed it with a stopwatch because I'm that kind of person) and then drop to 1.8GHz and stay there forever (thermal envelope they said but the heatsink was clearly inadequate) and your build times show exactly that pattern (the variance in your ten runs must be interesting too because I bet the first run after idle is faster and then they converge) and the power normalization is clever but I wonder about the accuracy

Nice work! I actually ran something similar on my home lab RK3588 board for a fun little project. Used a Noctua NH-L9a and it actually works — sustained 2.3GHz through the whole build. Your OVHcloud numbers make me think they're using the stock cooler or no cooler at all.

# my janky thermal fix
echo 75000 > /sys/class/thermal/thermal_zone0/trip_point_0_temp

Not recommending this for production obviously

The Hetzner scheduler weirdness — are you seeing cores parked on the same physical die? I had similar with their Graviton3 where NUMA topology was exposed wrong. Fixed with `taskset -c 0-3,8-11` to avoid the shared L3.

Would love to see your normalized perf/watt chart if you have one. My guess is RackNerd and Vultr are fighting for top spot.