Quick Answer: A gigabit adapter runs at 1000 Mbit/s of line rate, which appears as roughly 940 Mbit/s of payload once framing overhead is removed. The bus decides whether you see that figure or lose 10 to 15 percent of it. A PCIe card on a Gen2 x1 lane holds full duplex throughput at low host CPU load. A USB 3.2 Gen 1 dongle usually lands between 850 and 900 Mbit/s, and a USB 2.0 adapter cannot pass about 330 Mbit/s however the driver is tuned. Buy PCIe for fixed desktops, USB for laptops, and test the jack and cable before blaming the adapter.
Two adapters carrying the same gigabit label often differ by 300 Mbit/s on a real file transfer, and the gap sits in the interface, not in the PHY. Before ordering, decide whether the port is permanently installed, whether the machine will ever run jumbo frames, and what the processor is already doing. Those three answers remove most of the catalogue.
Where the 1000 Mbit/s goes
Gigabit Ethernet carries at most 1500 bytes of data behind 38 bytes of framing at minimum, before the preamble and the 12 byte inter-frame gap are counted. That arithmetic caps payload near 940 Mbit/s in one direction, so a healthy 1000BASE-T link tests at 930 to 945 Mbit/s and never at a clean 1000.
Full duplex doubles the aggregate to 1.88 Gbit/s because both directions run at once, but no single TCP stream will display that. One stream over a 20 ft (6 m) patch lead typically settles between 800 and 940 Mbit/s, and the variation comes from window size and the disk at each end rather than from the adapter.
The bus sets the ceiling
The PHY is the same part on almost every adapter in this class. What separates a 940 Mbit/s card from an 870 Mbit/s dongle is the path to the processor and the interrupt load that path creates.
PCIe x1 cards
A Gen2 x1 lane carries 500 MB/s in each direction, eight times what the PHY needs, and a Gen3 lane doubles that. The card sits on its own bus, so throughput stays flat while the USB stack is busy. In a virtualised host this is the difference between steady numbers and erratic ones.
USB 3.2 Gen 1 adapters
A USB 3.2 Gen 1 port signals at 5 Gbit/s, shared with protocol overhead and interrupt handling. Expect 850 to 900 Mbit/s one way, higher CPU cost per gigabyte and a small latency penalty under many small packets. A USB 2.0 adapter cannot pass about 330 Mbit/s of payload whatever the packaging claims.
What each build actually returns
These figures assume a single stream to a local server over Cat 6, standard 1500 byte frames, no jumbo frames and no offload tuning. Change those conditions and the payload moves by 20 to 40 Mbit/s, not by hundreds.
| Interface | Line rate | Payload seen | Host CPU | Best fit |
|---|
| PCIe 2.0 x1 | 1000 Mbit/s | 930-945 Mbit/s | Low | Fixed desktop |
| PCIe 3.0 x1 | 1000 Mbit/s | 930-945 Mbit/s | Low | Server, hypervisor |
| USB 3.2 Gen 1 | 1000 Mbit/s | 850-900 Mbit/s | Moderate | Laptop, dock |
| USB 2.0 | 480 Mbit/s | 280-330 Mbit/s | Moderate | Legacy only |
| M.2 PCIe | 1000 Mbit/s | 930-945 Mbit/s | Low | Small form factor |
Cable, jack and driver checks
The adapter cannot repair a poor channel. 1000BASE-T needs all four pairs, so a two-pair lead falls back to 100 Mbit/s and looks exactly like a dead card. A wire map test takes two minutes; replacing hardware costs a day. Any ethernet cable rated Cat 5e or above carries the 100 MHz bandwidth gigabit needs over the full 328 ft (100 m).
The jack matters less than the termination behind it. Gigabit shares the 8P8C socket with 100BASE-TX, and a port light turning amber rather than green usually means an open pair rather than a failed chip. Connector and protocol are separate things, which is the point of rj45 vs ethernet.
Driver settings shift the number without touching hardware. Leave interrupt moderation and receive side scaling on, and enable jumbo frames only where every switch and endpoint agrees on 9000 bytes. A slimline ethernet cable using 28 AWG conductors is fine at 6 ft (2 m) inside a rack and lossy at 50 ft (15 m) under 30 W of PoE.
Practical scenarios
- Situation: A 2015 desktop with a failing onboard PHY. Fit a PCIe 2.0 x1 card, disable the onboard port in firmware, and expect 940 Mbit/s with no change to the driver stack.
- Situation: A shared laptop dock feeding two desks. Choose a USB 3.2 Gen 1 adapter and budget for 880 Mbit/s, because the dock already shares its uplink with display traffic.
- Situation: A workshop machine 80 ft (24 m) from the switch, stuck at 100 Mbit/s. Re-terminate all four pairs at both ends first; two-pair termination is the usual cause.
- Situation: A virtualised host running four guests. Use PCIe 3.0 x1 and enable jumbo frames end to end; the gain shows up as CPU headroom, not as a higher speed test.
FAQ
Does a gigabit adapter need Cat 6?
No. 1000BASE-T runs over Cat 5e at 100 MHz across the full 328 ft (100 m). Cat 6 adds margin and supports 2.5GBASE-T and 5GBASE-T later, so buy it for the upgrade path rather than for gigabit itself.
Why does the port report 100 Mbit/s?
Almost always a wiring fault: one or more pairs open, a two-pair lead, or a crushed jacket near a staple. The PHY renegotiates downward instead of failing, so run a wire map before buying a replacement card.
Do jumbo frames raise throughput?
Only where every device in the path uses the same frame size, and even then the gain is CPU efficiency rather than headline rate. On a mixed network they cause fragmentation and retransmits that lower the number.
