2026-07-28 32 GB/s. That is the bidirectional bandwidth ceiling of a PCIe Gen3 x4 slot, the most common slot size for add-in network cards. A single 100GbE NIC can push 12.5 GB/s in one direction. Do the division and you get 2.5 seconds to fill a Gen3 x4 pipe at line rate. Now swap in a Gen4 x16 slot at 64 GB/s per direction and the picture changes entirely. The slot you plug your NIC into matters as much as the NIC itself.
Before comparing NICs, you need the raw numbers on PCIe lanes. Here is what each PCIe generation and slot combination actually delivers in one direction. Understanding PCIe lanes and their bandwidth is essential for matching the right slot to your NIC:
| PCIe Gen | Per-Lane Speed | x4 Total | x8 Total | x16 Total | |----------|---------------|----------|----------|-----------| | Gen3 | 985 MB/s | 3.94 GB/s | 7.88 GB/s | 15.76 GB/s | | Gen4 | 1.97 GB/s | 7.88 GB/s | 15.75 GB/s | 31.51 GB/s | | Gen5 | 3.94 GB/s | 15.75 GB/s | 31.50 GB/s | 63.01 GB/s |
These are theoretical maximums. Real-world throughput runs 3 to 5 percent lower due to encoding overhead and protocol headers. PCIe Gen3 uses 128b/130b encoding, so you lose roughly 1.5 percent to encoding alone. Gen4 and Gen5 use the same encoding scheme, so the overhead ratio stays consistent across generations (PCI-SIG, PCIe Base Specification Rev 5.0).
The key takeaway: a Gen3 x4 slot gives you roughly 3.8 GB/s of usable bandwidth across its PCIe lanes. A Gen4 x8 gives you about 15 GB/s across eight PCIe lanes. A Gen5 x16 gives you roughly 60 GB/s across all sixteen PCIe lanes. The gaps between these tiers are not incremental. They are step-function jumps that determine whether your NIC runs at full speed or hits a wall.
Here is where the math meets real hardware. A network card needs its PCIe slot to provide at least as much bandwidth as its port speed in each direction. Full-duplex traffic means a 25GbE NIC needs 25 Gbps (about 3.1 GB/s) in each direction simultaneously.
| NIC Speed | Full-Duplex Bandwidth | Minimum Slot (Gen3) | Minimum Slot (Gen4) | Comfortable Fit | |-----------|----------------------|---------------------|---------------------|-----------------| | 10GbE | 2.5 GB/s (each dir) | x4 (3.94 GB/s) | x4 (7.88 GB/s) | x4 Gen3 | | 25GbE | 6.25 GB/s (each dir) | x8 (7.88 GB/s) | x4 (7.88 GB/s) | x8 Gen3 | | 100GbE | 12.5 GB/s (each dir) | x16 (15.76 GB/s) | x8 (15.75 GB/s) | x8 Gen4 or x16 Gen3 | | 200GbE | 25 GB/s (each dir) | Not possible | x16 (31.51 GB/s) | x16 Gen4 | | 400GbE | 50 GB/s (each dir) | Not possible | Not possible | x16 Gen5 |
The uncomfortable truth: most server motherboards ship with one x16 slot and maybe one or two x8 slots, with the remaining slots wired as x4 or even x1. If you drop a 100GbE dual-port NIC into a Gen3 x8 slot, you are leaving 30 to 40 percent of your PCIe lane bandwidth on the table. The NIC can physically fit. The electrical lanes are there. But the slot is only wired for half the data paths the card expects.
A PCIe x4 slot is not a dead end. For 10GbE networking, it provides more than enough headroom. The LRES1016PF-SFP+, built on the Intel 82599 controller, is a single-port 10G SFP+ NIC designed for x4 slots. It delivers full 10GbE line rate through a Gen3 x4 connection with room to spare. In a server with limited high-lane-count slots, reserving x8 and x16 slots for cards that actually need them while running 10GbE management or out-of-band traffic through x4 slots is sound engineering.
The x4 slot also works for 10GbE copper. The LREC9811BT uses the Intel X550 controller to deliver 10GBASE-T over standard Cat6a cabling through a PCIe x4 interface. For environments where fiber is not practical, this keeps the slot cost low while delivering full 10G performance.
The rule is simple: if your NIC speed fits comfortably within x4 PCIe lane bandwidth with at least 20 percent headroom, x4 is the right choice. 10GbE at 1.25 GB/s per direction in a 3.94 GB/s Gen3 x4 slot uses only 32 percent of available PCIe lanes bandwidth. That is comfortable.
The 25GbE generation lands squarely in the x8 zone. A dual-port 25GbE NIC like the LRES1027PF-4SFP28 (Intel E810, 4x SFP28) requires PCIe x8 Gen3 or x4 Gen4 to run at full line rate across all four ports. In a Gen3 x8 slot, it gets 7.88 GB/s per direction. Four 25GbE ports at full duplex need 12.5 GB/s total bidirectional bandwidth, which a Gen3 x8 slot handles at 7.88 GB/s per direction (15.76 GB/s bidirectional). The math works.
The x8 slot is also where 100GbE single-port cards become viable on older server platforms. The LRES1019PF-QSFP28 uses the Intel E810 controller with a single QSFP28 port for 100GbE connectivity. In a Gen4 x8 slot, it gets 15.75 GB/s per direction, enough for full 100GbE line rate. In a Gen3 x8 slot, it gets 7.88 GB/s per direction, which caps throughput at roughly 63 Gbps per direction. For many workloads, especially when combined with jumbo frame MTU to reduce packet overhead, this is acceptable.
At 100GbE dual-port or 200GbE single-port, x16 is not optional. The LRES1014PF-2QSFP28 packs two 100GbE QSFP28 ports on an Intel E810 controller. Both ports at full line rate demand 25 GB/s bidirectional. A Gen3 x16 slot provides 15.76 GB/s per direction (31.52 GB/s bidirectional) across all sixteen PCIe lanes. A Gen4 x16 provides 31.51 GB/s per direction using the same PCIe lane count. For dual-port 100GbE with headroom, Gen4 x16 is the target.
For 200GbE and 400GbE adapters, PCIe Gen5 x16 is the only option that avoids bottlenecking. The LRES1260PF-QSFP112, based on the Broadcom BCM57608 controller, delivers 400GbE through a QSFP112 connector. At 50 GB/s per direction, it needs every lane of a Gen5 x16 slot to operate at full speed. Drop it into a Gen4 x16 slot and you cap at 31.51 GB/s per direction, leaving roughly 37 percent of the NIC's capability unused.
Here is the part that catches most people off guard. A physical x16 slot on your motherboard is not always wired as x16 electrically. Many server boards wire the second x16-length slot as x8 or even x4, sharing lanes with other slots or M.2 connectors. A NIC that physically fits in the slot will negotiate down to the electrical lane count.
Before buying a NIC, check three things:
1. The motherboard manual's slot specification table. Look for the electrical lane count, not just the physical connector size. A slot labeled "x16 (x4 mode)" gives you four lanes, not sixteen.
2. Lane sharing rules. Many boards share lanes between PCIe slots and NVMe M.2 slots. Populating an M.2 drive may cut your x16 slot down to x8. The motherboard manual lists these sharing configurations in a table usually buried in the specifications section.
3. CPU lane budget. The CPU provides a fixed number of PCIe lanes. On Intel Xeon Scalable (Sapphire Rapids), you get 80 PCIe Gen5 lanes from the CPU. On AMD EPYC (Genoa), you get 128 PCIe Gen5 lanes. If you populate multiple high-lane-count NICs, GPUs, or NVMe drives, you can exhaust the CPU's PCIe lane supply before you run out of physical slots.
For most server deployments in 2026, the practical recommendation breaks down by workload:
The pattern is consistent: match the PCIe lane bandwidth to the NIC speed with at least 15 to 20 percent headroom for protocol overhead. Anything less and you are paying for network bandwidth you cannot use.
Physically, no. An x8 card will not fit in an x4 slot unless the slot is open-ended (some server motherboards feature open-ended x4 slots that accept x8 cards). Electrically, even if it fits, the card will negotiate down to x4 speeds, which may bottleneck the NIC. Always check the motherboard manual for slot compatibility before attempting this.
Both matter, but in different ways. PCIe lane count determines your bandwidth ceiling. Generation determines the per-lane speed. A Gen4 x4 slot (7.88 GB/s) matches a Gen3 x8 slot (7.88 GB/s) in raw bandwidth. If your motherboard supports Gen4, you can use fewer PCIe lanes to achieve the same throughput. For NICs, this means a Gen4 x4 slot can fully support a 25GbE card that would need x8 on a Gen3 board.
On Linux, run `lspci -vvv -s
Yes. PCIe is fully backward and forward compatible at the physical and protocol level. A Gen5 NIC in a Gen4 slot will negotiate to Gen4 speeds and the available lane bandwidth of the Gen4 slot. You lose the Gen5 per-lane speed advantage, but the card functions correctly. The reverse also works: a Gen4 NIC in a Gen5 slot runs at Gen4 speeds.
