If you use a proxy client app (sing-box, Xray, Clash, Shadowrocket and friends), you have probably wondered whether your app choice matters. It turns out it matters a lot. With the same protocol and the same server, the slowest app I tested managed under 0.4 Gbit/s while the fastest passed 16 Gbit/s, and energy per GB differed by more than 40×.
The full interactive results are here: Tunnel Atlas. This post is the short version.
How I tested
The idea: take the network out of the picture so the only thing being measured is the app itself.
- Mac: MacBook Air M1, on battery, screen off, no app windows open.
-
Server on the same device: every app ran in TUN mode with the same Shadowsocks 2022 profile (
2022-blake3-aes-128-gcm) pointing at a sing-box server on the same Mac. iperf3 generated the traffic. - Speed: 5 s after a 2 s warm-up, over 1 and 4 connections, upload and download.
- Efficiency: a steady 200 Mbit/s download for 20 s, measuring the energy, CPU time and memory of the app's tunnel process.
- Fairness: a script ran the apps one at a time, two rounds, the second in reverse order. Every test checked its traffic really went through the tunnel, and any test with keyboard or trackpad input was thrown out and repeated.
- Score: 30% speed, 70% efficiency. Efficiency counts more because the top apps are already far faster than any home internet connection.
The sing-box command line is included as a baseline for "the core alone, no app around it".
Mac results (top 10 of 24)
| App | Score | Avg speed (Gbit/s) | Energy (J per GB) | Memory (MB) |
|---|---|---|---|---|
| Throne 1.4.0-beta.3 | 93.9 | 16.9 | 0.29 | 31 |
| sing-box CLI 1.15.0-alpha.12 | 82.4 | 16.3 | 0.34 | 29 |
| SFM 1.15.0-alpha.12 | 81.3 | 9.6 | 0.28 | 37 |
| Egern 2.20.0 | 64.2 | 3.9 | 0.26 | 7 |
| Surge 6.10.0 | 57.9 | 10.2 | 0.46 | 201 |
| Hiddify 5.0.0 | 49.1 | 9.6 | 0.57 | 191 |
| FlClash 0.9.1 | 48.4 | 10.5 | 0.61 | 26 |
| v2rayN 7.25.6 | 46.2 | 8.7 | 0.60 | 59 |
| Clash Verge Rev 2.5.8 | 45.7 | 7.7 | 0.58 | 27 |
| Koala Clash 1.5.0 | 45.6 | 10.6 | 0.66 | 25 |
At the bottom, several popular iOS-first apps running on the Mac (V2Box, Clash Plus, Streisand, incy, Happ) stayed under 0.7 Gbit/s and used 7–15 J per GB, roughly 25–55× the leaders.
iPhone results (6 apps)
Same idea on an iPhone 11 Pro: a test app on the phone runs the Shadowsocks 2022 server, so no Wi-Fi or internet is involved. Three rounds, run overnight with nobody touching the phone. Power is the whole phone's, minus the same test without a VPN.
| App | Score | Avg speed (Gbit/s) | Extra power at 200 Mbit/s (mW) |
|---|---|---|---|
| Egern 2.20.0 | 91.8 | 2.46 | 155 |
| Anywhere 2.4 | 63.0 | 3.27 | 300 |
| Shadowrocket 2.2.92 | 49.5 | 1.16 | 271 |
| Hiddify 4.0.0 | 16.8 | 0.34 | 756 |
| V2Box 10.1.9 | 13.2 | 0.36 | 979 |
| Happ 6.1.0 | 11.5 | 0.24 | 722 |
Egern is the most efficient and Anywhere is the fastest. The last three use about 5–6× the extra power of Egern for the same 200 Mbit/s.
Three things I didn't expect
1. MTU matters a lot. At the common MTU of 1500, an app handles about 2.7× more packets per GB than at 4000. Raising it from 1500 to 4000 lifted SFM's upload from 10.8 to about 14.3 Gbit/s, and Karing's speed from 0.9 to 2.4 Gbit/s (with its network stack also changed). On the Mac, the seven apps stuck at 1500 or below are all among the nine slowest. If your app lets you set the MTU, try 4000.
2. Same core, very different results. Many of these apps wrap the same open-source cores (sing-box or Xray), yet they land far apart. The app around the core (its TUN stack, MTU and settings) decides a lot.
3. Real use costs more than any lab. These tests use one proxy with no routing rules. In an earlier test through real servers, two chained proxies with routing rules used about 3× the energy per GB of a single proxy.
Caveats
This is independent, non-commercial testing on one Mac and one iPhone. Results depend on the device, OS and app version (versions as installed on 20–22 Oct 2026), and a real network adds its own costs. A few apps needed exceptions (for example, Hiddify 5.0.0 is a development build with a fixed MTU of 9000); they are all listed on the site.
More on the site
- Benchmark: every measure for all 24 apps (speed over 1 and 4 connections, energy, CPU, memory, iPhone idle cost).
- Routes: suggested 2026 setups for three situations (no censorship, a censored country, a shutdown or allowlist), with a diagram of who sees what along the path.
- Timeline: 58 VPN and proxy techniques from 1995 to 2026, each with its current status in four heavily censored countries.
- Sources: every reference behind the numbers.
The site source is on GitHub. If your favourite app is missing, or you think a result is off, open an issue and I'll look into it.
