A← Index

bare vs Chrome: a measured baseline

One real end-to-end test on a staging web app, run in Chrome and in bare, on the same machine and the same minute. This is the baseline that later work is measured against.

Date18 November 2026
barecommit 532cc4e9
Chrome155.0.8059.39
Runs20 of 20 passed
5.1xless memory (PSS)
1.3xless CPU time
1.25xfaster journey
1process, not 13

IThe test

The app is a production Next.js and React dashboard for a voice-AI product, on its staging server. The test does what a QA engineer does by hand:

  1. Log in.
  2. Create a client.
  3. Create an agent for it with a three-step wizard: setup, name and voice, review.
  4. Delete the agent.
  5. Delete the client.

Both browsers run the same step file. bare runs it with bare-drive. Chrome runs it through a small driver over the DevTools protocol, the same protocol that Playwright uses. The Chrome driver sends real mouse events and waits until an element stops moving before it clicks. After every run, a script checks the app and finds no test data left behind.

IIMemory

Memory is the clearest result. bare is one process. Chrome starts 13 processes for one tab: the browser, a GPU process, renderers, utilities and crash handlers.

Peak PSSChrome 5.12x
Chrome429 MiB
bare83.8 MiB
Peak RSS, all processesChrome 14.0x
Chrome1,216 MiB
bare86.6 MiB
cgroup memory.peakChrome 4.50x
Chrome353 MiB
bare78.6 MiB
ProcessesChrome 13.0x
Chrome13
bare1

Headless runs, mean of 6 each. PSS splits shared pages between processes, so it is the fair total. RSS counts shared pages once per process, so it overstates Chrome.

IIITime and CPU

bare finishes the journey 1.25x faster and uses 1.3x less CPU time. Most of the CPU gap is in the kernel. Chrome spends 9x more kernel time on starting processes and passing messages between them.

Journey timeChrome 1.25x
Chrome8.3 s
bare6.6 s
CPU time, totalChrome 1.30x
Chrome5.1 s
bare3.9 s
CPU time, kernelChrome 9.25x
Chrome0.9 s
bare0.1 s
CPU time, userChrome 1.10x
Chrome4.2 s
bare3.9 s
User instructionsbare 1.17x more
Chrome25.4 G
bare29.7 G

Where bare is slower. bare runs 17% more user-space instructions than Chrome. This app is mostly React and app scripts. Chrome compiles JavaScript with V8. bare walks the syntax tree and has no JIT. On a small static test page in an earlier comparison, bare used 250x fewer instructions. On a script-heavy app, the JavaScript engine is the gap to close.

IVWhere the time goes

Each phase of the journey, mean of the headless runs. Most of each phase is network and server time, so both browsers are close. bare is faster in every phase.

Log inChrome 1.17x
Chrome1.8 s
bare1.5 s
Create the clientChrome 1.37x
Chrome1.0 s
bare0.7 s
Create the agentChrome 1.10x
Chrome1.8 s
bare1.6 s
Delete the agentChrome 1.28x
Chrome1.9 s
bare1.5 s
Delete the clientChrome 1.44x
Chrome1.8 s
bare1.2 s

VAll numbers

Mean of the runs, with the spread as (max − min) / mean in brackets. Ratio is Chrome ÷ bare: above 1 means Chrome uses more.

Headless, 6 runs each
MetricChromebareRatio
Journey time (s)8.26 (14%)6.63 (13%)1.25x
CPU total (s)5.12 (3%)3.95 (2%)1.30x
CPU user (s)4.233.851.10x
CPU kernel (s)0.890.109.3x
User instructions (G)25.4 (2%)29.7 (2%)0.86x
Peak PSS (MiB)429 (3%)84 (2%)5.1x
Peak RSS sum (MiB)1,2178714.1x
memory.peak (MiB)353794.5x
Processes131
In a window, 4 runs each
MetricChromebareRatio
Journey time (s)7.96 (10%)6.84 (7%)1.16x
CPU total (s)5.51 (4%)4.45 (2%)1.24x
CPU user (s)4.304.360.99x
CPU kernel (s)1.210.1012.6x
User instructions (G)28.0 (5%)34.5 (3%)0.81x
Peak PSS (MiB)528 (1%)91 (5%)5.8x
Peak RSS sum (MiB)1,4039414.9x
memory.peak (MiB)398864.6x
Processes151

The Chrome driver is a separate Node process and is not in the Chrome columns. It added 0.28 s of CPU time and 76 MiB of memory to each run.

VIMethod and limits

One machine: AMD Ryzen 9 5900HX, 16 threads, Linux 7.2. Each browser ran in its own cgroup. CPU time and memory peaks come from the cgroup, user instructions from perf stat, and PSS from a sampler every 20 ms. Window runs used an isolated headless Wayland compositor, so no run touched the desktop.

Before the runs, I paused every other workload on the machine: other agents, 29 containers, a virtual machine, timers and file sync. Load average stayed between 0.3 and 0.9. The runs alternated between the two browsers, 6 pairs headless and 4 pairs in a window.

Not tested. A second machine, more runs, a real desktop window, a warm Chrome cache between runs (each Chrome run used a fresh profile), and Playwright itself. Wall times depend on the network and the staging server. The instruction counts are the most repeatable numbers.

SignedArjun