Fastest median visual load across five fresh browser runs.
Fire Pixel Host-Off · Measured 2 September 2026
Seven hosts. One identical website. Every file verified.
Cloudflare edge, GitHub Pages, two Ryzen shared hosts and three VPS routes served the same Fire Pixel page. The small fixture hid the difference. The image-and-font waterfall exposed it.
A dated Dublin snapshot, not a permanent global league table.
Cold LCP · lower is better
- Cloudflare Pages264 ms
- GitHub Pages272 ms
- HostXNow 9950X296 ms
- MechanicWeb 7950X304 ms
- CX43 through Tunnel304 ms
- CX23 HTTP/1.1344 ms
- CX43 direct388 ms
The first useful answer
The edge won. It did not destroy the 9950X.
Cloudflare led the full visual fixture, but HostXNow was only 32 ms behind on median cold LCP. Both scored 100 in the simulated mobile Lighthouse runs.
A measurable 32 ms laboratory gap, not a dramatic visitor-facing difference.
All timed visits returned 200, loaded the image and made zero external requests.
The same page, viewed three ways
Change the measure and the gap changes
Bars start at zero and show the measured value. Rank follows whether lower or higher is better.
Five fresh Chrome processes per route. Lower is better.
- 01 Cloudflare Static edge · HTTP/2 264 ms
- 02 GitHub Pages Static edge · HTTP/2 272 ms
- 03 HostXNow Shared LiteSpeed · HTTP/2 296 ms
- 04 MechanicWeb Shared LiteSpeed · HTTP/2 304 ms
- 05 CX43 Tunnel VPS through Tunnel · HTTP/2 304 ms
- 06 CX23 Small VPS · HTTP/1.1 344 ms
- 07 CX43 direct Direct VPS origin · HTTP/2 388 ms
The slow-end page load across six simultaneous isolated pages. Lower is better.
- 01 Cloudflare Static edge · HTTP/2 292 ms
- 02 GitHub Pages Static edge · HTTP/2 296 ms
- 03 MechanicWeb Shared LiteSpeed · HTTP/2 309 ms
- 04 CX43 Tunnel VPS through Tunnel · HTTP/2 335 ms
- 05 CX23 Small VPS · HTTP/1.1 361 ms
- 06 HostXNow Shared LiteSpeed · HTTP/2 377 ms
- 07 CX43 direct Direct VPS origin · HTTP/2 399 ms
Median completed pages per second across three six-page trials. Higher is better.
- 01 Cloudflare Static edge · HTTP/2 10.79
- 02 GitHub Pages Static edge · HTTP/2 10.36
- 03 CX43 Tunnel VPS through Tunnel · HTTP/2 10.03
- 04 MechanicWeb Shared LiteSpeed · HTTP/2 9.90
- 05 CX23 Small VPS · HTTP/1.1 9.17
- 06 HostXNow Shared LiteSpeed · HTTP/2 8.98
- 07 CX43 direct Direct VPS origin · HTTP/2 8.80
Full visual fixture
The complete measured snapshot
A normal desktop visit made nine requests and transferred about 103 to 106 KiB on the compressed HTTPS mirrors. Every one of the 19 source files matched its manifest.
| Route | TTFB | Cold LCP | Load | Pages/s | p95 load | Mobile |
|---|---|---|---|---|---|---|
| Cloudflare Pages Static edge · HTTP/2 | 108 ms | 264 ms | 208 ms | 10.79 | 292 ms | 100 / 1.40 s |
| GitHub Pages Static edge · HTTP/2 | 123 ms | 272 ms | 223 ms | 10.36 | 296 ms | 100 / 1.36 s |
| HostXNow Ryzen 9 9950X Shared LiteSpeed · HTTP/2 | 125 ms | 296 ms | 266 ms | 8.98 | 377 ms | 100 / 1.42 s |
| MechanicWeb Ryzen 9 7950X Shared LiteSpeed · HTTP/2 | 122 ms | 304 ms | 276 ms | 9.90 | 309 ms | 100 / 1.40 s |
| Hetzner CX43 through Cloudflare Tunnel VPS through Tunnel · HTTP/2 | 168 ms | 304 ms | 241 ms | 10.03 | 335 ms | 100 / 1.47 s |
| Hetzner CX23 direct HTTP Small VPS · HTTP/1.1 | 111 ms | 344 ms | 293 ms | 9.17 | 361 ms | 94 / 1.99 s |
| Hetzner CX43 direct origin Direct VPS origin · HTTP/2 | 167 ms | 388 ms | 328 ms | 8.80 | 399 ms | 100 / 1.48 s |
The CX23 route is HTTP/1.1 without equivalent public TLS, so it is diagnostic context rather than a like-for-like HTTPS ranking. HostXNow was measured the following day with the identical method.
Method before marketing
What had to match before a timer started
Byte-for-byte fixtures
All 19 source files were fetched from every mirror and checked against the same SHA-256 manifest.
Fresh cold loads
Each sample launched a new Chrome process and context. Shared CDN caches were left as found.
Six pages together
Three trials loaded six simultaneous pages in isolated contexts and recorded throughput, p95 and failures.
What the CPU badge missed
Configuration moved more than processor size
- HTTP/2 changed the burst ranking. An older HTTP/1.1 tool produced the opposite conclusion, so that result was excluded.
- The Tunnel beat the direct route. The exact same CX43 backend was faster through Cloudflare on cold LCP and concurrent p95.
- Compression changed the visual result. The CX23 had a strong TTFB but transferred about 36 KiB more and painted later.
- Third-party scripts outweighed hosting. On the live homepage, roughly 78 per cent of transferred bytes came from analytics, consent and measurement services.
The real homepage
Hosting was not its largest payload
The complete live Fire Pixel waterfall contained 25 requests and about 506 KB. Only nine requests and 109 KB were first-party. Once the origin is already fast, consent and measurement choices can outweigh another server upgrade.
The audit also found a real defect: late-loading font weights caused desktop layout shift. Preloading the exact weights took CLS from about 0.07 to 0.00 across the confirmation runs.
Open the evidence
Every visual fixture is still live
The mirrors are deliberately noindex and point their canonical attribution back to Fire Pixel.
Cloudflare
The quickest cold visual load and the steadiest six-page burst in this Dublin snapshot.
GitHub Pages
Nearly level with Cloudflare on the full fixture and the six-page burst.
HostXNow
Only 32 ms behind Cloudflare on median cold LCP. The processor label did not decide the result.
MechanicWeb
Effectively tied with HostXNow visually, with a stronger p95 in the simultaneous-page test.
CX43 Tunnel
The Tunnel route beat the direct route to the exact same server. Removing it was not an automatic win.
CX23
Shown for diagnosis, not a like-for-like TLS ranking. HTTP/1.1 and compression outweighed the CPU size.
CX43 direct
The same backend as the Tunnel route, but slower on cold LCP and p95 in this network path.
Limits
What this test cannot honestly claim
One Dublin client cannot establish a global winner. A modest burst cannot establish maximum capacity. Shared-host results can move with neighbouring tenants, and lab Lighthouse is not field data.
The useful claim is narrower: these routes served the same verified page, under the published method, on the published date. We will add locations and dynamic PHP tests rather than stretching the first result beyond its evidence.
Questions and challenges
Host-Off FAQ
Which host won the first Host-Off?
Cloudflare Pages led the median cold LCP and concurrent pages per second from Dublin. GitHub Pages was close behind. That does not make either platform universally fastest: location, cache state, protocol, workload and shared-host contention can change the order.
Was the Ryzen 9 9950X host slow?
No. HostXNow delivered the full visual fixture with a 296 ms median cold LCP and a 100 mobile Lighthouse performance score. Cloudflare was 32 ms ahead on cold LCP, which is measurable in a lab but not a dramatic visitor-facing difference.
Did the Cloudflare Tunnel make the VPS slower?
Not in this test. The Tunnel route and direct route used the exact same CX43 backend, yet the Tunnel route was 84 ms faster on median cold LCP and 64 ms better at p95 under the six-page trial.
Can these results identify the fastest web host?
No. This is a dated, reproducible snapshot from one client location. It compares these routes with these fixtures. It is useful evidence about delivery configuration, not a permanent global league table.
Can another provider join the Host-Off?
Yes. A provider can host the same public fixture. It must remain unmodified, tracking-free and noindex, and every manifest file must match before its measurements are published.
From benchmark to service
Want the right route for your site, rather than the loudest specification?
Fire Pixel Fast Hosting & Care places static sites at the edge and managed WordPress sites on an isolated LiteSpeed account, then measures the deployed result.