CloudFlare vs Vercel benchmark

This repo is meant to benchmark SSR performance between CloudFlare and Vercel. There's a lot of misinformation going around and I'm annoyed. So I wrote a really annoying component and server rendered it in a bunch of different ways.

Results

From my findings, Vercel is 1.2x to 5x faster than CloudFlare for server rendering.

I also noticed some CF runs would perform TERRIBLY on Next and SvelteKit. Like 1/5th of attempts taking 10+ seconds (mean is closer to 1.2s). Horrible variability.

Do these numbers really matter? Meh. Not really. Most "slowness" for web apps is APIs and queries being slow. CPU is rarely the bottleneck.

Regardless, here's the results laid out (you can see individual runs in the results directory).

I ran 100 iterations with every framework/provider combo. The results will shock you!

All Vercel tests were run with "performance" function environments (2 vcpu and 4gb ram). I ran multiple runs on "Standard" as well (included in /results), difference wasn't as large as expected (~10% hit overall).

next-js

Platform	Mean	Min	Max	Variability
Cloudflare	1.895s	0.800s	3.971s	3.171s
Vercel	0.534s	0.343s	1.442s	1.098s

Winner: Vercel (3.55x faster)

react-ssr-bench

Platform	Mean	Min	Max	Variability
Cloudflare	0.476s	0.227s	1.383s	1.156s
Vercel	0.138s	0.059s	0.635s	0.576s

Winner: Vercel (3.45x faster)

sveltekit

Platform	Mean	Min	Max	Variability
Cloudflare	0.292s	0.078s	1.038s	0.960s
Vercel	0.113s	0.058s	0.552s	0.494s

Winner: Vercel (2.59x faster)

shitty-sine-bench

Platform	Mean	Min	Max	Variability
Cloudflare	23.401s	8.122s	46.326s	38.204s
Vercel	36.025s	30.376s	38.156s	7.781s

Winner: Cloudflare (1.54x faster)

realistic-math-bench

Platform	Mean	Min	Max	Variability
Cloudflare	1.469s	0.751s	3.387s	2.636s
Vercel	0.702s	0.463s	1.136s	0.673s

Winner: Vercel (2.09x faster)

vanilla-slower

Platform	Mean	Min	Max	Variability
Cloudflare	0.220s	0.104s	0.620s	0.516s
Vercel	0.208s	0.119s	0.743s	0.624s

Winner: Vercel (1.06x faster)

FAQ

Why do this?

Someone was lying on the internet and it went viral. I was annoyed. This should show why.

Isn't this measuring round trip time? Why not measure actual compute time?

Fun fact - you _can't use performance.now() on CloudFlare. It's frozen due to spectre side channeling. Since CloudFlare runs your code in an isolate instead of a vm, you're sharing memory with other workers by other devs. They have to do shit like this to keep your data safe.

What is the "shitty-sine-bench" test about?

There's another youtuber who claimed CF is 3x faster than Vercel. He's mostly wrong.

CF is ~2x faster at computing floating point numbers and sine/cosine operations. In more realistic math, Vercel is faster (see realistic-math-bench).