PCIe Lane Sharing Explained: How Expansion Cards and M.2 SSDs Can Limit GPU Bandwidth

XDA Developers writer Hamlin Rozario hit this on his MSI MPG X870E Carbon WiFi, an AM5 board. The same problem showed up on his older Asus ROG Strix X570-E. Below is what the vendor documentation says, what the benchmarks say, and how to check your own board before you spend money.

What the MSI manual actually says​

The X870E Carbon WiFi has three full-length slots, and they are not equal. PCI_E1 comes from the CPU and supports up to PCIe 5.0 x16. PCI_E2 also comes from the CPU and supports up to PCIe 5.0 x4. PCI_E3 comes from the X870E chipset and supports up to PCIe 4.0 x4.

The manual also notes that PCI_E1, PCI_E2 and M2_2 share the bandwidth. The exact combinations are in a configuration table in the manual.

MSI's specification page spells out the CPU dependence:

• With Ryzen 9000/7000 processors, PCI_E1 and PCI_E2 run as PCIe 5.0 x16/x0 or x8/x4.
• With Ryzen 8700/8600/8400 processors, they run as PCIe 4.0 x8/x0.
• With Ryzen 8500/8300 processors, they run as PCIe 4.0 x4/x0.

So the behavior is not universal across every CPU that fits the socket. A reviewer at ALKTech noted that with a Ryzen 8000 APU, the primary slot drops to PCIe 4.0 x8 or x4 and the second slot is disabled entirely.

The trade-off in practice​

Rozario's account matches MSI's documentation. If you populate PCI_E2 or M.2_2, the CPU lanes split into x8/x4/x4, and his RTX 4090 runs at PCIe 4.0 x8 instead of x16. Populating the chipset-fed third slot avoids that.

MSI's lane-sharing table lists the third slot as sharing bandwidth with nothing in that CPU group. PCI_E3 shows "No" under shared bandwidth, while PCI_E1, PCI_E2 and M2_2 all share with each other.

One nuance matters here. The trigger is not just "a card in slot two." An independent reviewer at Real Hardware Reviews found that even a PCIe 3.0 x1 add-in card in the middle slot drops the top slot to x8, and so does an older PCIe 3.0 x2 M.2 drive in M.2_2. The slow card does not use the lanes, but the split still happens.

In short:

• Leave PCI_E2 and M.2_2 empty and the GPU slot keeps x16.
• Use either one and the GPU slot drops to x8.
• Use PCI_E3 (and the chipset M.2 sockets) for add-in cards if you want to keep x16.

Full-length does not mean x16​

PCIe generation sets the speed per lane, and lane count multiplies it. At the same generation, x8 is half the theoretical bandwidth of x16.

GamersNexus makes the same point in its PCIe scaling article. On this board the second slot is only x4 despite its length. For slot wiring in general, GamersNexus explains that a slot can look x16 in length while being wired for fewer lanes. Its tip is to look at the pins inside the slot: half populated means x8, and fewer means x4. Chipset-fed slots may also run an older PCIe generation and add a hop back to the CPU.

For most add-in cards that is fine. Capture cards and network adapters usually need no more than x4. Multi-drive M.2 adapters are a different case.

Multi-SSD adapters need extra care​

A four-drive card like the Asus Hyper M.2 x16 needs lanes and board support for splitting them (bifurcation). In an x4 slot you will not get four working drives. ASUS's compatibility FAQ stresses that the maximum number of supported SSDs varies by CPU and motherboard. It also gives a lookup procedure: find your board on the model list, check your CPU, and check which slot the card goes in. It adds that filling PCIEX16_2 with four SSDs can leave no lanes to map drives on PCIEX16_3.

The ASUS FAQ also lists card generations. The Hyper M.2 x16 Gen 5 card supports PCIe 4.0 and 5.0. On W790 boards with a Gen 4-only slot, ASUS says to set that slot's speed to Gen 4 in the BIOS.

The ASUS X570-E example in the packet is firmware-dependent. The manual describes a UEFI setting that switches the second slot between standard x8 and a RAID mode that splits it into x4+x4. It also warns that installing other devices while in RAID mode can prevent booting. I could not open that manual in this session, so treat that detail as secondhand until you read your own manual.

How much does x8 really hurt?​

Less than the headline suggests. GamersNexus tested an RTX 5090 at PCIe 5.0, 4.0 and 3.0, all at x16. Its summary was a roughly 1–4% maximum difference between Gen 5 and Gen 3. PCIe 3.0 x16 matches PCIe 4.0 x8 in theoretical bandwidth, so that result is a decent proxy for x8 on a PCIe 4.0 slot.

The caveats are real:

• The tests covered 16 scenarios on a Ryzen 7 9800X3D bench. GamersNexus says it did not test machine learning or AI workloads, so those results don't apply there.
• It compared generations at x16 and did not measure x8 directly.
• GamersNexus flags a different case: cards that are natively x4 or x8, such as the RX 6500 XT, can suffer much more in an older slot.
• Rozario's own caveat is that an older PCIe 3.0 board is a different story. Dropping a GPU from 3.0 x16 to x8 could hurt more.

Note that there is no free lunch. Even if frame rates barely change, you are giving up capability you paid for. A PCIe 5.0 GPU that falls back to x8 is only at PCIe 4.0 x16 equivalent bandwidth.

A pre-purchase checklist​

1. Identify your exact motherboard model and CPU. On boards like this one, the CPU changes the lane map.
2. Open the manual's expansion-slot specs and look for asterisks on the slot and M.2 lines.
3. For each slot, note whether it is CPU- or chipset-connected, its maximum PCIe generation, and its lane width.
4. Find the lane-sharing or configuration table and identify which M.2 sockets or slots share lanes with your GPU slot.
5. Check for BIOS/UEFI options that change slot modes or bifurcation.
6. Read the add-in card's requirements, especially multi-drive adapters.
7. Plan your slots so the busiest devices sit on non-shared connections.

The takeaway​

Sharing lanes between the GPU slot, a second slot and an M.2 socket is a design choice. It is not a defect. Boards do it to offer more connectors than the CPU has lanes. Forum users have noted that many X870E boards share lanes this way, partly because AMD requires USB4 lanes to come from the CPU. That is a user's explanation, not an official statement, so weigh it accordingly.

For a modern GPU on a current board, the cost is usually small. The real risk is spending money on an expansion card and then learning your board can't run it fully. A few minutes in the manual costs far less than that.