RunsOn m9gd.xlarge · local NVMe vs Blacksmith 4 vCPU arm64
RunsOn does each build for 5.9–7.1× less and 1.7–2.3× faster. Both 4 vCPU arm64.
All three builds 6.3× cheaper on RunsOn: $0.0140 vs $0.0876RunsOn 1.8× faster: 8m 56s vs 16m 26s
Rust build 7.1× cheaper on RunsOn: $0.0029 vs $0.0210RunsOn 2.2× faster: 1m 47s vs 3m 59s
TypeScript build 6.5× cheaper on RunsOn: $0.0017 vs $0.0111RunsOn 2.3× faster: 50s vs 1m 55s
Docker build 5.9× cheaper on RunsOn: $0.0094 vs $0.0555RunsOn 1.7× faster: 6m 19s vs 10m 32s
RunsOn m9gd.xlarge · local NVMe
jobs ran from Amazon.com (US)
Blacksmith 4 vCPU
jobs ran from Secured Servers (US)
Side by side
Every measurement both runners share, cost first. The dot marks the better value; gap is the larger over the smaller.
| Measurement | RunsOn | Blacksmith | Gap |
|---|---|---|---|
| Cost | |||
| Cost per Rust build | $0.0029 | $0.0210 | 7.1× RunsOn |
| Cost per TypeScript build | $0.0017 | $0.0111 | 6.5× RunsOn |
| Cost per Docker build | $0.0094 | $0.0555 | 5.9× RunsOn |
| Price per minute | $0.0014 | $0.0050 | 3.7× RunsOn |
| Real builds, wall time | |||
| Rust build (spotify-player) | 1m 47s | 3m 59s | 2.2× RunsOn |
| TypeScript build (opencode) | 50s | 1m 55s | 2.3× RunsOn |
| Docker build (PostHog) | 6m 19s | 10m 32s | 1.7× RunsOn |
| CPU & memory | |||
| PassMark single-thread (points) | 2,474 | 1,353 | 1.8× RunsOn |
| PassMark CPU Mark, all cores (points) | 5,630 | 3,783 | 1.5× RunsOn |
| sysbench CPU, 1 thread (events/s) | 3,767 | 929 | 4.1× RunsOn |
| 7-Zip compression (MIPS) | 41,586 | 17,447 | 2.4× RunsOn |
| Memory read bandwidth | 28.1 GiB/s | 9.0 GiB/s | 3.1× RunsOn |
| Measurement | RunsOn | Blacksmith | Gap |
|---|---|---|---|
| Workspace disk (fio) | |||
| Random read, 4k QD32 (IOPS) | 87,221 | 253,175 | 2.9× Blacksmith |
| Random write, 4k QD32 (IOPS) | 43,591 | 35,622 | 1.2× RunsOn |
| Random read, 4k QD1 (IOPS) | 13,375 | 32,740 | 2.4× Blacksmith |
| QD1 read latency, p99 | 111 µs | 37 µs | 3.0× Blacksmith |
| Sequential read | 585 MiB/s | 1.7 GiB/s | 2.9× Blacksmith |
| Sequential write | 281 MiB/s | 829 MiB/s | 3.0× Blacksmith |
| Sync write, 4k + fdatasync (IOPS) | 14,860 | 9,366 | 1.6× RunsOn |
| Network, cache & Docker | |||
| Download from GitHub | 147 MiB/s | 100 MiB/s | 1.5× RunsOn |
| Cache restore, 4 GB (actions/cache) | 23s | 21s | 1.1× Blacksmith |
| Cache save, 4 GB (actions/cache) | 19s | 14s | 1.4× Blacksmith |
| docker pull | 4.7s | 12.9s | 2.8× RunsOn |
| Database | |||
| pgbench read/write (TPS) | 8,444 | 2,948 | 2.9× RunsOn |
| pgbench read-only (TPS) | 65,158 | 24,073 | 2.7× RunsOn |
| Queue | |||
| Queue under burst load | 22s | 12s | 1.8× Blacksmith |
- ≈ even: under 5% apart, or a build inside both runners' run-to-run spread
- Queue: identical jobs queued at the same instant (burst suite)
- RunsOn at EC2 spot, its default (7-day average of the cheapest us-east-1 zone's spot price, fetched 16 Nov 2026); 19 of its jobs fell back to on-demand (AWS had no spot capacity for the type) and are priced at the spot rate over the same billed time
- Cost: a managed SaaS runner against RunsOn self-hosted on your AWS account; the effort of running it was not measured