Why Is Your GPU Running at PCIe x8 or x4 Instead of x16? Causes, Checks & Fixes

2026-10-18 ｜ GPU Repair ｜ 14 min ｜ HK GPU Fix Editorial

Key Takeaways

• A GPU-Z reading of x8 or x4 is not automatically a fault. The most common cause of a low reading is an idle screenshot: the link drops to a lower generation at idle to save power and renegotiates up under load.
• Always read the link state under load. In GPU-Z, click the question mark next to Bus Interface and run the render test, then read the value while the test is running. An idle reading proves nothing either way.
• Physical slot size does not equal electrical lane count. A full-length-looking slot may only be wired x8 or x4 electrically, and that is by design, not damage.
• Lane sharing is per-motherboard. Populating a second full-length slot, or a lane-sharing M.2 slot on some boards, can drop the top slot to x8. Your board manual's shares bandwidth footnotes are the authority for your board.
• Some cards are natively x8, such as the RTX 5060 Ti with native PCI-Express 5.0 x8. On such a card, x8 at full load is correct behaviour.
• A genuine fault looks different: the link stays below the card's and board's rated maximum under sustained load, or it comes with crashes, black screens, or frequent link retraining.
• On gaming impact, the only hard numbers available are reported figures from secondary summaries of TechPowerUp's RTX 4090 test: at 1080p, Gen 4 x16 to x8 cost about 2 percent, dropping to x4 cost another about 4 percent, and Gen 3 x4 cost about 21 percent. Those figures apply to that test and that card only.

How PCIe Negotiation Works

A PCIe link is a point-to-point serial connection made of 1 to 16 lanes. Packet data is striped across lanes, and peak throughput scales with lane count. Lane count is automatically negotiated during device initialisation and can be restricted by either endpoint. A single-lane (x1) card in a multi-lane slot auto-negotiates the highest mutually supported lane count. The link can dynamically down-configure to fewer lanes, providing failure tolerance if bad or unreliable lanes are present. The standard defines link widths x1, x2, x4, x8, and x16.

Link training is handled by the LTSSM (Link Training and Status State Machine), which moves through Detect, Polling, Configuration, and then L0, the fully operational state. During training, the LTSSM negotiates link width and link speed. If a link becomes unstable, the LTSSM enters the Recovery state to resynchronise clocks or downtrain the speed to maintain a stable connection.

Negotiation picks the highest value both ends support. A Gen 5 card in a Gen 3 slot runs at Gen 3 speed; the newer card will function, but communication speeds are limited to the older side's bandwidth. PCIe is backwards compatible across generations, and PCI-SIG confirms that newer PCIe products interoperate and maintain backwards compatibility with all previous generations.

Two structural facts explain most confusion. First, physical slot size does not equal electrical lane count: a full-length-looking slot may only be wired x8 or x4 electrically. Second, bifurcation splits one x16 slot or link into smaller groups, for example x8 plus x8, or four x4 links, and it only works when the CPU and BIOS both support the split.

Bandwidth: What x8 and x4 Actually Mean

The table below gives usable throughput in one direction, from the PCI Express link-performance table. Do not mix these one-direction figures with PCI-SIG's headline bidirectional totals (for example, 256 GB/s for PCIe 6.0 x16) without labelling the direction.

Because each generation doubles per-lane throughput, several combinations are arithmetically equal in theoretical one-direction bandwidth: Gen 4 x8 equals Gen 3 x16 at 15.754 GB/s; Gen 5 x8 equals Gen 4 x16 at 31.508 GB/s; Gen 3 x8 equals Gen 2 x16 at 7.877 GB/s; and Gen 4 x4 equals Gen 3 x8 at 7.877 GB/s. This is why a width drop on a newer generation is often less costly than it looks, and why a width drop on an older generation can hurt much more.

Reading GPU-Z Correctly

GPU-Z on Windows is the standard tool for checking the negotiated link state. Its Bus Interface field shows two values: the maximum the card supports, then the currently negotiated state. For example, PCIe x16 3.0 @ x8 1.1 means the card supports x16 at Gen 3, and is currently running x8 at Gen 1.1.

GPU-Z has a built-in render test: click the question mark next to Bus Interface, then Start Render Test. It puts GPU usage at roughly 99 to 100 percent so the link leaves its idle power state and the true negotiated maximum can be read. One widely cited rule: if the reading changes from, for example, x8 1.1 to x16 3.0 under load, the link is fine. If it changes to x8 3.0, or stays at x16 1.1 only, there is an issue to investigate.

On Linux, the same comparison is available from lspci -vv, which shows LnkCap (capability) versus LnkSta (current negotiated status), for example LnkCap Speed 8GT/s, Width x8 versus LnkSta Speed 8GT/s, Width x8. The nvidia-smi command can query link generation and width, current versus maximum. The dmidecode -t slot output confirms the physical slot capability as a complementary check.

The Idle-Downshift Trap

At low GPU load, the PCIe bus runs at a lower speed to save power; as load rises, it switches to the high-bandwidth state. Normally, at low loads, the PCIe x16 bus runs at PCIe x16 1.1, and as the load increases it goes to a high-bandwidth state, generally PCIe x16 3.0 on a Gen 3 platform. Modern NVIDIA GPUs dynamically downshift the PCIe link to Gen 1 when the device is idle to save power. As soon as a workload hits the GPU, the link renegotiates up to its trained maximum generation. Seeing Gen 1 at idle on a Gen 5 slot is expected behaviour, not a fault.

The mechanism behind this is ASPM (Active State Power Management): the PCIe serial link is powered down when there is no traffic. The low-power modes L0s (one direction) and L1 (whole link including reference clock) trade power for exit latency. Windows exposes this as PCI Express Link State Power Management (Off, Moderate, or Maximum power savings) in Power Options advanced settings.

One nuance must be stated honestly: the well-documented idle behaviour is a generation drop, for example x16 3.0 to x16 1.1. Whether width also drops at idle, for example x16 to x8 or x2, varies by platform and GPU. Forum users report it, but no authoritative source consulted states it as a general rule. So the correct check is: read the link state under load. An idle reading proves nothing either way. Do not claim idle always drops width, and do not claim idle never drops width.

Genuine Causes: By Design

Slot population and lane sharing

On most motherboards, the first PCIe x16 slot runs at maximum PCIe bandwidth, x16, while the second full-length slot runs at x8 speed only. Manufacturer manuals say this explicitly. The GIGABYTE Z890 AI TOP manual states that the PCIEX8 slot shares bandwidth with the PCIEX16 slot, and when the PCIEX8 slot is populated, the PCIEX16 slot operates at up to x8 mode, while the chipset slot runs at x4. The same pattern appears on the X570 AORUS ULTRA: the PCIEX8 slot shares bandwidth with the PCIEX16 slot, and when the PCIEX8 slot is populated, the PCIEX16 slot operates at up to x8 mode. That board's documentation adds: for optimum performance, if only one PCI Express graphics card is to be installed, be sure to install it in the PCIEX16 slot.

M.2 slots can share lanes too, but the rules are per-board. On some Intel LGA1700 motherboards, the x16 PCIe 5.0 link reserved for graphics is split to provide support for a PCIe 5.0 M.2 NVMe SSD. Running at x12 is not possible, so the graphics card is left with eight lanes only. But leaving the M.2 Gen 5 slot empty and installing the SSD in the CPU-attached Gen 4 M.2 slot will not touch GPU connectivity, so the card can run at x16. On AMD Socket AM5 motherboards, the M.2 NVMe Gen 5 slots are wired to the processor, and installing drives on these slots does not subtract any lanes from the x16 PEG slot. One board-specific user report illustrates the same idea: on an ASUS ROG Crosshair X870E Hero, installing a second PCIe 5.0 drive in M.2_2 or M.2_3 drops PCIEX16_1 to x8, while M.2_4 and M.2_5 run at PCIe 4.0 and do not affect the GPU slot. That report is board-specific and illustrative only, not a general rule.

CPU lane budgets frame the same picture. A secondary source reports that Ryzen 7000 and 9000 and Intel Core Ultra 200s desktop CPUs can process up to 24 lanes (20 PCIe 5.0 direct to CPU plus 4 PCIe 4.0 via the chipset link), while Intel 13th and 14th generation CPUs offer 16 PCIe 5.0 plus 8 PCIe 4.0, with the 16 shared between the GPU slot and a CPU-attached M.2 slot, so populating that M.2 slot drops the GPU to x8. The same source adds the essential caution: not every motherboard has this issue, and you need to check the manual.

Multi-GPU setups share lanes by the same logic: two cards in SLI or CrossFire will run at PCIe x8 each, because the CPU PCIe x16 lanes are shared equally between the two cards.

The writer guidance here is blunt: lane-sharing rules are per-motherboard. The only reliable source for a given board is that board's manual and spec page, especially the shares bandwidth footnotes. Never state a universal rule like any M.2 drops the GPU to x8.

The card itself may be natively x8

Some GPUs are physically and electrically x8 cards. For example, the RTX 5060 Ti is described by TechPowerUp as having native PCI-Express 5.0 x8. Such a card will show @ x8 even at full load, and that is correct behaviour. Always check the card's own specification (maximum link width) before concluding anything is wrong. GPU-Z's first number is the card's maximum.

A chipset-wired lower slot

Bottom full-length slots are often wired through the chipset at x4 electrically. A card installed there will run at x4 by design. Moving it to the top slot, usually the CPU-connected x16 slot, restores full width. A chipset-wired x4 slot is not a broken x16 slot.

An older board or CPU

A Gen drop caused by a board or CPU that only supports that generation is not a fault. A Gen 5 card in a Gen 3 slot runs at Gen 3 speed, and that is the negotiated maximum for that combination.

BIOS and Firmware Settings

BIOS can force a PCIe generation. TechPowerUp used controls in the UEFI BIOS setup program to limit Gen 5 to Gen 4 to Gen 3 for scaling tests. Some boards have a BIOS option to switch a slot between modes or adjust lane behaviour, and some let you switch a slot between SATA and PCIe mode. A BIOS update can also fix link-negotiation bugs: if the motherboard BIOS is outdated, updating it may resolve the problem.

Check for a forced generation first, then for a bifurcation or slot-split setting that divides the x16 link, then for any unusual lane-sharing option on your board. Trying Auto for generation is the safe default unless you are using a riser cable rated for a lower generation, in which case the generation should be forced to the riser's rating (see below).

One dated, board-specific quirk is worth a single sentence: on some older Gigabyte boards, a Turbo USB 3.0 BIOS option was reported to consume CPU PCIe lanes and leave the graphics card at PCIe x8 only. Treat it as an example of check unusual BIOS options on your board, not as a common cause.

Physical Causes: Seating, Pins, and Riser Cables

Seating and dirty pins

Reseating the card is a genuine fix for a real mechanism. Remove the graphics card, clean the PCIe connector contacts with isopropyl alcohol, and put it back properly. Sometimes dust or even part of a warranty sticker can be present on the PCIe connector. Clean the slot with an air duster. If the data pins required to enable a full PCIe x16 interface are blocked, the graphics card will work at a lower bandwidth interface, preferably PCIe x8 or lower. Clean with contact cleaner or 99.9 percent isopropyl alcohol.

CPU and socket pins are a possible but less common cause: if one of the pins is dirty and not making contact, then all of the CPU PCIe lanes may not be available for the graphics card, and this can cause the GPU to run at lower PCIe x8 speed. CPU reseating is high-effort, so treat it as a later step, not a first step.

The link's ability to down-configure on bad lanes connects these facts: a persistent x8 or x4 after reseating, especially with instability, is consistent with a lane-level signal problem such as a dirty or damaged pin, slot damage, or poor contact. That is an inference from the down-configuration mechanism and the pin-blocking evidence, not a directly sourced diagnostic claim.

Riser cables

Riser cables are a common physical cause of link trouble in vertical-mount builds, but the established facts are about generation and stability, not width. Mixing a PCIe 4.0 GPU and motherboard with a PCIe 3.0 riser cable often requires manual BIOS throttling to display an image. If you use a PCIe 3.0 riser with a PCIe 4.0 GPU and motherboard, the system will attempt to communicate at Gen 4 speeds and fail, so you must manually set the motherboard slot to Gen 3 in the BIOS first. Cheap risers rated for Gen 3 often cause crashes or black screens at Gen 4 and Gen 5 speeds, and the standard fix forces a lower generation in the BIOS, which quietly trades bandwidth for stability.

The decisive test is simple: temporarily test the GPU without the riser cable, directly in the slot, to rule the riser out. One user report described a Gen 4-rated riser (case-bundled) limiting a card to x8 even with the generation forced in BIOS, with the card running x16 without the riser. Treat that as anecdotal: a riser can limit width in an individual case, but no controlled test was consulted proving that risers generally limit width. Do not claim riser cables generally limit width to x8.

Card or board fault

Cross-testing separates the two. If the card still runs x8 in another system, the card is the likely culprit. If the card runs x16 in another PC, the board or slot is suspect. The same logic applies in reverse for the motherboard.

One claim to avoid: a single source says a missing or faulty 6 or 8-pin PCIe power connector can in some cases leave the card working but at much lower speeds. That is a weak, single-source claim, and the same source admits the PC usually will not boot at all in that situation. This article does not present it as an established cause of x8. Checking that power connectors are fully seated is still a sensible generic step, without claiming it changes link width.

How Much Does It Matter?

The only hard numbers available come from one test, and they must be attributed honestly: reported figures from secondary summaries of TechPowerUp's test, which used an RTX 4090 with an i9-13900K on a Z790 board. TechPowerUp's own article page returned 403 to the researchers, so the numbers below come via secondary summaries and are presented as such, not as directly verified from TechPowerUp.

At 1080p, moving from Gen 4 x16 to Gen 4 x8 cost about 2 percent performance; dropping to Gen 4 x4 cost another about 4 percent, or about 6 percent total against x16; and Gen 3 x4 (equivalent to Gen 4 x2) cost about 21 percent, described as clearly unacceptable. At 1440p and 4K, the Gen 4 differences did not change that much, and Gen 3 x4 was slightly better at 4K. The summary's practical reading was that as long as the card is running on PCIe Gen 4 x4 or above, performance should be okay. Another summary of the same test described the RTX 4090's loss in PCIe 4.0 x8 mode as a negligible, inconsequential amount of performance, and noted that the difference between Gen 4 x16 and Gen 3 x16 was almost negligible too.

Do not generalise those exact percentages to all cards. The results are for an RTX 4090, a very high-end card, and a specific test suite. One single-author site estimated a tiny performance loss of around 1 to 2 percent at most for a GTX 1080 Ti-class card at Gen 3 x8, with significant loss at Gen 1.1 x8, Gen 2.0 x8, or Gen 3.0 x4. Present that only as one author's view, not as established fact.

The qualitative picture has support: PCIe GPU bottleneck gaming fears are mostly overblown for full x16 cards, because moving a modern GPU from Gen 4 to Gen 5 changes average frame rates a few percent at most. The exceptions are real, though. Cards with reduced lane counts on older boards, and games that overflow video memory, both lose measurable performance over a slow link. In many real-world gaming scenarios, x8 on newer PCIe generations can still perform very well, but it depends on the GPU and workload.

Workloads that stress the link more are the ones that move large amounts of data between CPU or system memory and the GPU: games that overflow VRAM and spill to system memory, GPU compute or AI work streaming data, and video capture or throughput tasks. Cards with smaller VRAM put more stress on the link when the card has to dip into system memory. Do not claim x8 never matters, and do not claim x4 is fine. Both are unsupported.

Diagnostic Order

Follow this order. Stop as soon as a step resolves the question.

1. Check under load first. Open GPU-Z, note Bus Interface, click the question mark, start the render test, and read the value while the test runs. If it reaches the card's rated maximum, for example x16 4.0, stop: nothing is wrong. Do not diagnose from an idle screenshot.
2. Check the card's own maximum. GPU-Z's first number, and the card's spec page: a native-x8 card showing x8 is correct.
3. Check the slot. Is the card in the top full-length slot, usually the CPU-connected x16? Is anything in the second full-length slot or a lane-sharing M.2 slot? Read your motherboard manual's shares bandwidth footnotes. They are the authority for your board.
4. Check BIOS. Is the PCIe generation forced to an old generation? Is a bifurcation or slot-split setting enabled? Is an unusual lane-sharing option set? Try Auto for generation, or force the generation your riser cable is rated for if using a riser.
5. Check physical. Power off, reseat the card, inspect the gold fingers and slot for dust, debris, or sticker residue, and clean with isopropyl alcohol and an air duster. If using a riser cable, test the card directly in the slot without it.
6. Isolate. Cross-test: the card in another PC (still x8 points to the card), another card in this slot (still x8 points to the board or slot). Check for a BIOS update from the board maker.
7. Decide. If it stays at x8 on a Gen 4 or newer platform and games and workloads run normally, the practical impact is usually small, but a persistent x4, a Gen 1 or Gen 2 link under load, crashes, or black screens point to a real fault needing inspection or repair. If a card needs professional inspection, the published reference range is HK$400-2,800 total; actual fee set by the post-inspection quote.

Normal Behaviour vs Real Fault

Common Misdiagnoses

1. Diagnosing from an idle GPU-Z screenshot. This is the most common mistake.
2. Treating x8 as a fault when the card is natively x8.
3. Treating by-design lane sharing (second slot or M.2) as a fault, when the manual says it will happen.
4. Treating a chipset-wired x4 slot as a broken x16 slot.
5. Treating a Gen drop caused by an older board or CPU as a fault.
6. Assuming a riser cable just works at any generation. A Gen 3 riser in a Gen 4 system can black-screen unless the BIOS is set to Gen 3.
7. Assuming reseating cannot help. Blocked pins genuinely can drop width, but equally, reseating will not fix by-design sharing.
8. Confusing the two numbers in GPU-Z (PCIe x16 4.0 @ x8 1.1). The first is the maximum, the second is the current state. Readers who read only the second number panic for nothing.

For related diagnostics, see the companion monitoring guide (article ID 108) for reading GPU-Z and HWiNFO64, the no-detection guide (article ID 106) if the card is not detected at all, the display-output guide (article ID 105) for no signal over HDMI or DisplayPort, and the VBIOS guide (article ID 104) for firmware-side issues.

FAQ

My GPU-Z shows x8 1.1. Is my card broken?

Almost certainly not on that evidence alone. x8 1.1 is a typical idle reading. Run the GPU-Z render test by clicking the question mark next to Bus Interface and read the value while the test runs. If it rises to the card's rated maximum, the link is fine. Only a below-maximum reading under sustained load is worth investigating.

My card runs at x8 even under load. Should I worry?

First check whether that is by design. Is the card natively x8, like the RTX 5060 Ti with native PCIe 5.0 x8? Is anything in the second full-length slot or a lane-sharing M.2 slot? Is the card in a chipset-wired lower slot rather than the top slot? Is the BIOS forcing an old generation or a bifurcation split? Work through the diagnostic order above before concluding anything is faulty. On a Gen 4 or newer platform, x8 alone usually has a small practical impact in gaming, but a persistent x4 or a Gen 1 or Gen 2 link under load points to a real problem.

Does x8 hurt gaming performance?

The only hard numbers available are reported figures from secondary summaries of TechPowerUp's RTX 4090 test: at 1080p, Gen 4 x16 to x8 cost about 2 percent, dropping to x4 cost another about 4 percent, and Gen 3 x4 cost about 21 percent. Those figures apply to that test and that card only. The qualitative picture is that x8 on newer generations usually performs well, while reduced lane counts on older boards and games that overflow VRAM lose measurable performance over a slow link.

Can a riser cable cause x8?

The established riser facts are about generation and stability: a Gen 3 riser in a Gen 4 system can fail or black-screen unless the BIOS slot is set to Gen 3, and cheap risers cause crashes or black screens at Gen 4 and Gen 5 speeds. One user report described a Gen 4-rated riser limiting a card to x8, but no controlled test was consulted proving risers generally limit width. The decisive test is to run the card directly in the slot without the riser.

My card used to run x16 and now runs x8. What changed?

A new symptom with no configuration change points toward physical degradation: seating, dirty or damaged pins, slot damage, or a failing component. Work through the diagnostic order: check under load, check BIOS, reseat and clean the connector, test without the riser if you use one, and cross-test the card in another PC and another card in this slot to isolate card versus board.

Should I force the PCIe generation in BIOS?

Usually leave it on Auto. Forcing a generation is appropriate when using a riser cable rated for a lower generation (set the slot to the riser's generation), or when diagnosing a negotiation problem step by step. A forced old generation left in place by accident is itself a common cause of a persistent low reading, so check it.

Sources: PCI-SIG PCIe 6.0 specification and FAQ; Wikipedia PCI Express link-performance table; Vectree PCIe protocol engineering notes; GraphicsCardHub PCIe troubleshooting guide; Baaz GPU power-management analysis; TechPowerUp RTX 5060 Ti PCIe x8 scaling review; TechPowerUp RTX 4090 PCIe scaling test as reported by Club386 and Epicdope; GIGABYTE Z890 AI TOP and X570 AORUS ULTRA manuals; XDA Developers bifurcation guide; NVIDIA enterprise support PCIe configuration notes; MakeUseOf lane-budget guide.