Wi-Fi speeds, the long version
Top line
Measured on a proper multi-stream speed test from the same desk, the old machine went from about 16–21 Mbit/s to 244 Mbit/s down and 228 Mbit/s up. The new card alone didn't do that: three separate problems were stacked on top of each other, and each one hid the next.
| When | What changed | Result |
|---|---|---|
| before | old Atheros card, 2.4 GHz, Wi-Fi 4 | roughly 9–19 Mbit/s |
| day 1 | new MT7927 card working (driver, regulatory domain, MAC fixes) | 130 down / 107 up, 3 ms ping (the author's own speed test) |
| a week later | "internet is very slow" again | 13–19 Mbit/s. The card had fallen back to 2.4 GHz |
| same day | the Wi-Fi extender turned out to be a zombie and was unplugged | 5 GHz finally held: 44.6 / 50.5 Mbit/s, about 3× |
| that evening | pinned to 5 GHz (160 MHz wide), OFDMA/MU-MIMO turned on in the router | 244 down / 228 up |
The zombie extender
An old extender in the house still accepted connections but forwarded nothing, and it broadcast on the same 5 GHz channel as the router. Every time the new card tried 5 GHz it landed on the extender's copy of that channel, got a weak signal and died. That made it look as if the card was too weak for 5 GHz, and the notes recorded exactly that conclusion. Unplugging the extender proved it wrong: the router took the channel back and the card held it.
The notes correct themselves
The most honest part of the record is how often it overturns itself. The notes contain three numbered corrections to their own conclusions, including:
- "The internet connection is the ceiling, ~15–28 Mbit/s." This was wrong. A game download running in the background was eating the line during the measurements. A later note reads: "Every throughput and latency number taken during it was garbage."
- "The card's 5 GHz is too weak to hold a link." This was mostly the zombie extender, not the card.
- A rule that came out of it: before any throughput or latency measurement, check the link is actually idle for 5–10 seconds.
Measurement traps found along the way:
- A single download from one distant server understates real capacity: about 64 Mbit/s on the same link where a parallel speed test read 244.
- One public speed-test endpoint starts returning one-byte error pages after repeated use. That produces absurd numbers like 1,739 Mbit/s on a 20 Mbit/s line.
It didn't stay fixed
A month later the same machine, now a server, measured 6.2 Mbit/s on a clean single download, and it kept dropping off the network entirely. The likely cause is the card's antennas. NetworkManager made it worse by never retrying after a drop. That led to a small watchdog service that checks the network the way a LAN client sees it and reconnects after two failed checks. The real fix is an ethernet cable.