Benchmarking a cloud computer with V8

Benchmarking a cloud computer with V8

How we ran the V8 Web Tooling Benchmark on a 4 vCPU homemade computer, what it scored, what it says about CPU speed, and what it doesn't.

20 Nov 20264 min readhomemade team

How fast is a cloud computer at the work developers actually do? Boot time tells you when you can start. It doesn't tell you how long tsc, a bundler or a test run takes once you're in. For that you need a CPU benchmark, and for JavaScript work the most useful one we know is the V8 Web Tooling Benchmark.

This post covers what that benchmark measures, exactly how we ran it on a homemade computer, what it scored, and what the score does and doesn't tell you.

The result

On 19 November 2026, a homemade computer with 4 vCPU and 8 GB of RAM scored a median of 27.33 runs per second over five runs.

RunScore (runs/s)
127.33
226.48
326.94
427.60
527.85
Median27.33

Real speeds vary with load.

What the benchmark measures

The V8 Web Tooling Benchmark is published by the V8 project. It runs the core of the JavaScript tools web developers use all day, on real inputs: Babel, TypeScript, Acorn, Prettier, PostCSS, Terser, the CoffeeScript compiler and more. Its README says the goal is to measure only JavaScript performance, not I/O. Each tool runs repeatedly and the benchmark reports a score in runs per second. The headline number is the geometric mean across all of them. Higher is better.

It suits a dev environment for two reasons:

  • It is the workload. Parsing, transforming and printing code is what your editor, your bundler and your type checker do. A synthetic integer benchmark can look great on a CPU that is slow at this.
  • It is single-threaded. It measures how fast one core runs JavaScript. That is what decides how long most build steps take, since many of them don't spread across cores.

The second point also means it says nothing about many cores: a computer with more vCPUs does not score higher on it. More vCPUs help with parallel work, like running tests in parallel, compiling several packages at once or running several agents at once.

How we ran it

We wanted a method anyone can repeat on their own computer, on ours, or on someone else's. So we used the benchmark as published, with no changes.

git clone --depth 1 https://github.com/v8/web-tooling-benchmark wtb
cd wtb
npm install
node dist/cli.js

The score is the "Geometric mean" line at the end of the output. We ran it five times in a row on a fresh computer and report the median, plus the lowest and highest run.

A few details that matter more than you'd expect:

  • Node version. The repo pins Node 10 in its .nvmrc. Our computers install the pinned version automatically when you enter a folder that asks for it, so a plain node in that folder is Node 10. On an earlier test, Node 10 scored 10.55 where Node 26 scored about 13 on the same computer. Newer V8 is much faster here. We ran every reported number with the computer's default Node, v26.13.1, by calling it by its full path.
  • One computer at a time. We ran the benchmark with no other test computers running next to it. Under load you will see lower and noisier numbers.
  • Fresh and warm runs. On older tests we saw the first runs on a brand new computer come out a little lower than later ones. Five runs and a median smooth that out.

Comparing with other providers

StarSling publishes an open sandbox benchmark that includes a Node.js web tooling score for many providers at a 4 vCPU / 8 GiB target. Their run from 16 November 2026 lists, for example, Namespace at 23.86, Daytona (VM) at 21.7 and E2B at 14.29 runs per second. The leaderboard has the full table and their method.

We are not on that leaderboard, and our 27.33 does not come from their harness. They run the benchmark through the Phoronix Test Suite with Node 22, on each provider's sandbox at that size. We ran the benchmark directly with Node 26. Different Node versions alone move this score a lot. So treat the numbers side by side as a rough guide, and run the commands above yourself if the difference matters to you. Our comparison pages quote their published numbers next to ours with the same caveat.

What it doesn't tell you

A CPU benchmark is one number. It leaves out a lot of what makes a dev environment feel fast:

  • Disk. Installing dependencies is often disk-bound, not CPU-bound.
  • Network. Cloning a repo or pulling packages depends on the path to the registry.
  • Many cores. Parallel test runners and monorepo builds use them, and this test doesn't.
  • Start time. How long until you can type at all. We wrote about that in how we got click to terminal under 4 seconds.
  • Noisy neighbours. Computers share capacity, so a busy moment can slow one down. We share CPU fairly between computers, but a score from a quiet moment is the best case.

Try it on your own computer

If you have a homemade computer, the commands above work as is. Run node --version inside the wtb folder first: if it says v10, the repo's pin is active, so call a newer Node by its full path. Then compare the geometric mean with your laptop.56 for us, with other work running.

See pricing for the sizes, and the CLI for opening a computer from your terminal.

Try homemade

Every project gets its own cloud computers, with Claude Code and Codex signed in. The Hobby plan is free.

Related posts