Protocol choice decides three things you will actually feel: how much CPU your tunnel burns, how fast it reconnects after the network changes, and how much throughput survives on your hardware. Everything else in the feature list is secondary to those three.
A single handshake based on Curve25519, ChaCha20-Poly1305 and BLAKE2s, running over UDP. Stateless by design: the server keeps no session table, so a peer that changes address simply sends the next packet and continues. The whole implementation is a few thousand lines, which is why it costs so little CPU and why the roaming behaviour is good.
TLS-based, available over UDP or TCP, with a wide choice of cipher suites and a long history of interoperability. TCP mode survives hostile networks better than raw UDP, at the cost of head-of-line blocking and noticeably more overhead. Session setup is heavier: certificate exchange plus a TLS handshake, which is visible as a slower reconnect.
Native to iOS and macOS and well supported on Windows and Android, which makes it the least invasive choice on mobile. Its MOBIKE extension handles address changes cleanly, so hand-offs between Wi-Fi and cellular are fast. Configuration is the fiddly part, and the cipher negotiation has more moving pieces than WireGuard.
| WireGuard | OpenVPN | IKEv2/IPsec | |
|---|---|---|---|
| Transport | UDP only | UDP or TCP | UDP (500/4500) |
| Handshake | One round trip | TLS, several round trips | Two round trips (IKE_SA + CHILD_SA) |
| CPU cost | Lowest | Highest | Low to moderate |
| Reconnect after network change | Immediate | Full handshake | Fast via MOBIKE |
| Mobile support | Third-party app | Third-party app | Built into OS |
| Battery impact on phones | Moderate (app wake-ups) | Higher | Lowest (system stack) |
| Cipher flexibility | Fixed, modern set | Configurable legacy and modern | Configurable |
Published figures are almost always measured on server-grade hardware, which is not the machine you own. On consumer devices the practical ceilings look roughly like this, and they move with CPU generation:
| Device class | WireGuard | OpenVPN (UDP) | IKEv2 |
|---|---|---|---|
| Modern laptop, 8+ cores | 800 Mbps – multi-gigabit | 250–500 Mbps | 500–900 Mbps |
| Older laptop, 2–4 cores | 250–450 Mbps | 80–180 Mbps | 200–350 Mbps |
| Phone, mid-range | 150–350 Mbps | 60–120 Mbps | 150–300 Mbps |
| Travel router | 80–250 Mbps | 25–60 Mbps | 60–150 Mbps |
These are ranges, not promises. Measure your own: the same line, same server, both protocols, five runs each, median of the last four. If WireGuard is not at least twice the OpenVPN figure on the same device, the bottleneck is elsewhere — usually MTU or a single-core server.
Most "protocol is slow" reports resolve to one of these four parameters:
| You want | Pick | Why |
|---|---|---|
| Maximum throughput on a fast line | WireGuard | Lowest per-packet cost, kernel implementations available |
| Fastest reconnect on a phone that moves between networks | IKEv2 | MOBIKE plus the system stack, no app involved |
| Compatibility with old devices or restrictive networks | OpenVPN (TCP) | Widest support, survives UDP-hostile paths |
| Simple self-hosting | WireGuard | Two peers, one interface, no certificate authority |
| Least battery drain on iOS | IKEv2 | Handled by the OS networking stack |
ping the same host in each state; record the added latency.