Quick answer
How do I benchmark a VPS with YABS or Geekbench 6?
To benchmark a VPS, run curl -sL yabs.sh | bash: YABS (version v2026-10-31 on 2 November 2026) returns disk (fio), network (iperf3) and Geekbench 6 results in one pass, and ./geekbench6 --upload runs Geekbench 6 alone. Geekbench 6 scores a Core i7-12700 at 2,500, and double the score means double the speed. Cobaltfell quotes no score of its own yet; its dated results follow its first test machine.
What does YABS measure?
YABS (Yet Another Bench Script) runs three tests in one go: disk with fio, network with iperf3, and CPU with Geekbench 6. The version we read on 2 November 2026 was v2026-10-31.
- Disk (fio): four mixed random tests, half reads and half writes, at block sizes of 4k, 64k, 512k and 1m. Each runs for 30 seconds with direct I/O, an I/O depth of 64 and 2 jobs, on a 2 GB test file (512 MB on Arm and RISC-V) written in the directory you start from.
- Network (iperf3): download and upload against public iperf3 servers, with 8 parallel streams. The README warns that it tries to fill the port for about 20 seconds per location, 10 in each direction.
- CPU (Geekbench): Geekbench 6 by default. Since 24 July 2026, the
-7flag runs Geekbench 7 instead;-4and-5run the older versions.
It prints the system details first: processor model, vCPU count, RAM, disk, kernel and, unless you skip it, the network owner of the IP address.
How do I run YABS?
The one-line form downloads the script and runs it; the safer form lets you read it first. The README itself asks you to use it at your own risk, and its fio and iperf3 binaries are built by GitHub Actions, with checksums published.
curl -sL yabs.sh | bashcurl -sL yabs.sh -o yabs.sh
less yabs.sh
bash yabs.shcurl -sL yabs.sh | bash -s -- -i -w yabs-$(date +%F).jsonWhich YABS flags will you use most?
| Flag | Effect |
|---|---|
-i | Skip the iperf3 network test |
-r | Test fewer iperf3 locations, to use less bandwidth |
-f or -d | Skip the fio disk test |
-g | Skip Geekbench |
-n | Skip the lookup of network information |
-5 / -7 | Run Geekbench 5 or 7 instead of 6 |
-j | Print the results as JSON |
-w <file> | Write the JSON results to a file |
How do I run Geekbench 6 on its own?
Download the Linux build from Primate Labs, unpack it and run geekbench6. The version below, 6.7.2, is the one YABS used on 2 November 2026; take the current number from geekbench.com if it has moved on. Primate Labs lists Linux support from Ubuntu 18.04, on x86-64 and AArch64.
mkdir -p ~/gb6 && cd ~/gb6
curl -sL https://cdn.geekbench.com/Geekbench-6.7.2-Linux.tar.gz | tar xz --strip-components=1
./geekbench6 --uploadThe run ends with a Single-Core Score, a Multi-Core Score and a link to the full result on the Geekbench Browser. On Arm servers, the file name ends in LinuxARMPreview instead of Linux. A licence key is optional: YABS reads one from a geekbench.license file if you have it.
How do I read the results?
Read each number for what it measures, and compare it only with results from the same tool and version.
| Result | What it tells you | What it does not |
|---|---|---|
| Geekbench 6 single-core | Speed of one thread: page renders, request handling, most of a compile step. 2,500 = a Dell Precision 3460 with a Core i7-12700 | Whether that speed holds at 3 a.m. and at 9 p.m. |
| Geekbench 6 multi-core | How the vCPU work together on one shared task | A per-vCPU figure: it does not grow in step with the vCPU count |
| fio 4k | Small random reads and writes, the pattern of databases and package installs | Sequential speed for large files |
| fio 1m | Large-block throughput: backups, media, big copies | Latency of small requests |
| iperf3 | Throughput to a few public test servers at that moment | Your route to your own users; busy servers are skipped |
Three facts from Primate Labs' Geekbench 6 documentation help with reading the CPU scores:
- The scale is linear. The baseline is 2,500, and double the score means double the performance.
- Integer work weighs more. Each composite score weights integer workloads at 65% and floating-point at 35%.
- Multi-core uses shared tasks. Geekbench 6 splits one larger task across threads instead of running one copy per thread as Geekbench 5 did, so its multi-core score scales less than the vCPU count. Compare a 16-vCPU result with another 16-vCPU result.
Scores from Geekbench 5, 6 and 7 are not interchangeable: Geekbench 7, announced on 23 July 2026, brings new workloads. Keep the version beside every number you quote.
How do I make a benchmark fair?
One run tells you how the VPS behaved for ten minutes; three runs at different hours tell you how it behaves. According to Cobaltfell's survey of 2 November 2026, 10 of the 22 no-KYC crypto hosts that call a CPU dedicated named no processor precise enough to date it, so on those plans lscpu and a benchmark are the only way to learn what you rent. A short routine:
- Run on an idle VPS. YABS loads every vCPU, the disk and the network port; do it before production traffic arrives.
- Repeat at three hours of the day, for example morning, evening and night. On a dedicated vCPU, the Geekbench scores should stay close; a wide spread points to contention.
- Log steal time during the run. Keep
mpstat -P ALL 5open in a second session; what CPU steal time is and how to check it explains the reading. - Write down the context: date,
lscpumodel name, vCPU count, kernel, YABS or Geekbench version, city. - Mind your traffic. The iperf3 test pushes the port for about 20 seconds per location; use
-ror-ion a metered plan.
A dedicated vCPU should show the same behaviour at every hour. The page on the difference between a shared vCPU and a dedicated vCPU explains why a shared one often does not, and which is the best dedicated CPU VPS provider in 2026 lists the hosts whose promises are worth testing.
A timed job of your own completes the picture: an encode, as in can I transcode video with FFmpeg on a VPS without a GPU, or a model run, as in can I run Ollama on a VPS without a GPU.
Where are Cobaltfell's own results?
Not published yet, on purpose. Our own YABS and Geekbench 6 results, together with the steal time we measure, will be published with their date once our first test machine is delivered. Until then we quote no performance figure of our own.
When they appear, each result will carry its date, the plan, the city, the processor that lscpu reports and the version of the tool, beside the steal time logged during the run. Until then, any number you see for Cobaltfell elsewhere did not come from us.
You can run the same tests on your own VPS from the first login: Root over SSH. Paste a public key at checkout and password login stays off; leave the field empty and the team sets a temporary root password, shown in your dashboard, that you replace at first login. Plans and prices are on how much a dedicated-core VPS costs at Cobaltfell.
What other questions do people ask about VPS benchmarks?
Is it safe to pipe YABS into bash?
It runs code from the internet as your user, so read it first if you can: download it with curl -sL yabs.sh -o yabs.sh, look through it, then run bash yabs.sh. Its README says to use it at your own risk and publishes checksums for its binaries.
How long does YABS take?
The four fio tests take about two minutes, Geekbench a few more, and the network test about 20 seconds per iperf3 location. The sample run in the README finished in 12 minutes 49 seconds; -i makes it much shorter.
Do I need a Geekbench licence?
No. YABS and the plain geekbench6 --upload command run without one and upload the result to the Geekbench Browser. A licence key unlocks extra features; YABS reads it from a geekbench.license file.
Why is my multi-core score not the single-core score times my vCPU count?
Because Geekbench 6 measures threads working together on one task, which costs coordination. Its documentation says this "shared task" model may not scale as well as the separate copies Geekbench 5 ran per thread.
Can I compare a Geekbench 6 score with a Geekbench 5 or 7 score?
No. Each version has its own workloads and scaling. Compare Geekbench 6 with Geekbench 6 only, and quote the version with the number.
Will benchmarking use my traffic allowance?
The iperf3 part does: it fills your port for about 20 seconds per location, in both directions. Use -r for fewer locations or -i to skip it. The disk and CPU tests use no traffic beyond downloading the tools.
Sources (6)
- Yet Another Bench Script (YABS), README — masonr on GitHub (latest entry 31 Oct 2026, read 2026-11-02)
- yabs.sh, version v2026-10-31 — masonr on GitHub (read 2026-11-02)
- Geekbench 6 Benchmark Internals — Primate Labs (May 2024, read 2026-11-02)
- Geekbench 7 announcement — Primate Labs (23 July 2026, read 2026-11-02)
- Geekbench 6.7.2 for Linux (download file) — Primate Labs (available 2026-11-02)
- mpstat(1) manual page — man7.org (sysstat) (checked 2026-11-02)