Geekbench 6 · fio · iperf3 · 40 runs each
Benchmarks
Published in full, including the runs that make us look worse than a competitor. The methodology is at the bottom and the script is a single curl.
01 — Head to head
Single-core, Geekbench 6
Geekbench 6 single-core, median of 40 runs on production nodes.
The comparison rows are deliberately unnamed. They are our own measurements on comparable instance types bought at list price, and publishing a competitor’s name against a number we measured ourselves — on hardware we do not control, at a moment we chose — would be dishonest. The methodology is below; run it against whoever you are considering and compare like for like.
Single-core, Geekbench 6 · higher is better
Storage latency, p99, milliseconds · lower is better
02 — Every size
Multi-core, Geekbench 6
| Plan | vCPU | Single | Multi | Multi-core | 4K IOPS | Storage | Seq read |
|---|---|---|---|---|---|---|---|
| PULSE P1 | 2 | 2180 | 4,120 | 186,000 | 5,100 MB/s | ||
| PULSE P2 | 4 | 2180 | 7,940 | 224,000 | 5,600 MB/s | ||
| PULSE P3 | 6 | 2180 | 11,480 | 268,000 | 6,100 MB/s | ||
| PULSE P4 | 8 | 2180 | 14,900 | 312,000 | 6,400 MB/s | ||
| TORQUE T1 | 4 | 1720 | 9,800 | 402,000 | 11,200 MB/s | ||
| TORQUE T2 | 8 | 1720 | 18,900 | 512,000 | 12,400 MB/s | ||
| TORQUE T3 | 16 | 1720 | 35,600 | 684,000 | 13,800 MB/s | ||
| TORQUE T4 | 32 | 1720 | 68,200 | 902,000 | 14,600 MB/s | ||
| NITRO N1 | 4 | 3480 | 13,100 | 596,000 | 12,900 MB/s | ||
| NITRO N2 | 8 | 3480 | 25,400 | 742,000 | 13,900 MB/s | ||
| NITRO N3 | 12 | 3480 | 36,800 | 884,000 | 14,400 MB/s | ||
| NITRO N4 | 16 | 3480 | 44,100 | 1,020,000 | 14,900 MB/s | ||
| FORGE F1 | 4 | 2260 | 11,900 | 498,000 | 12,100 MB/s | ||
| FORGE F2 | 8 | 2260 | 22,800 | 646,000 | 13,200 MB/s | ||
| FORGE F3 | 16 | 2260 | 43,600 | 812,000 | 14,100 MB/s | ||
| FORGE F4 | 24 | 2260 | 62,400 | 1,140,000 | 14,800 MB/s | ||
| VAULT V1 | 4 | 1840 | 8,600 | 62,000 | 2,400 MB/s | ||
| VAULT V2 | 6 | 1840 | 11,400 | 78,000 | 2,800 MB/s | ||
| VAULT V3 | 8 | 1840 | 14,800 | 94,000 | 3,100 MB/s | ||
| VAULT V4 | 12 | 1840 | 26,900 | 118,000 | 3,600 MB/s | ||
| FUSION G1 · RTX 4000 Ada | 8 | 1780 | 18,400 | 620,000 | 13,100 MB/s | ||
| FUSION G2 · RTX 4090 24 GB | 12 | 1780 | 27,600 | 780,000 | 13,900 MB/s | ||
| FUSION G3 · L40S 48 GB | 16 | 1780 | 36,200 | 940,000 | 14,600 MB/s | ||
| FUSION G4 · H100 80 GB NVL | 24 | 1780 | 52,800 | 1,180,000 | 14,900 MB/s |
03 — Methodology
Methodology
Forty runs per configuration, spread over seven days, on production nodes carrying real customer load. Median reported, not best. Storage figures come from fio with a 4K random-read profile, queue depth 32, sixteen jobs, on a fully provisioned volume. Network figures are from iperf3 between two of our own PoPs and from RIPE Atlas probes for third-party paths.
Reproduce it yourself
On any CrestVPS instance, run the command below. It prints the same table you see above for your own machine.
The script writes nothing outside /tmp and uninstalls itself when it finishes.
04 — Silicon