DatacenterDisk
Server compatibility and memory

Why does my multi-drive NVMe card show only one drive, or none?

Sources checked · page updated · the question as people ask it: “I put a multi-drive NVMe adapter card (an x16 card with four M.2 sockets, an x8 card with two M.2 or two U.2 drives) in a PCIe slot and only the first drive shows up, or none do - is the slot not bifurcated, is the slot not wired for it, or do I need a different card, and how do I tell which?”

Quick answer

A passive multi-drive card gives each drive its own x4 share of the slot's lanes, so the slot has to be split in the firmware: Supermicro documents that a two-drive card in a slot left at x8 or x16 shows one drive, at x2x2x2x2 shows none, and shows both only at x4x4 (x4x4x4x4 in an x16 slot). If the split is set and drives are still missing, the slot may be wired narrower than its connector, share its lanes with an M.2 socket or another slot, hang off the chipset, or sit on a CPU that cannot make that split; count the NVMe controllers in lspci and read the board's slot table to tell which. A new card is the fix only when no slot on the board can be split, and then only a card with an on-board PCIe switch that divides the lanes itself.

What are you seeing?

Pick the closest symptom to open the matching diagnostic path. This is a guide to what to check and in what order; it does not read your system and does not diagnose it.

Count what the bus sees and work out what the card is
  1. 1.Start by counting what the host enumerated. Every NVMe drive reports PCI class 0108, 'Non-Volatile memory controller' (prog-if 02, NVM Express), whatever its name string says, so counting that class counts the drives visible on the host's PCI bus. Compare the number with the drives on the card.
    lspci -nn | grep -c '\[0108\]'
    ReadingA number below the drives installed means the missing ones are not visible on the host's PCI bus: the slot is not split for a passive card, the slot is wired narrower than the card needs, its lanes went to another slot or an M.2 socket, or the board or CPU cannot make the split (p2 and p3). On an Intel platform with VMD enabled, drives behind VMD can also be missing here: run the VMD check on our U.2 detection page, linked below; one reply in a ServeTheHome thread reports its best results with VMD disabled. A number equal to the drives installed means card and slot have done their job and the fault sits above the PCI bus, in the driver or a passthrough binding (p4).
  2. 2.Print the bus as a tree to see what kind of card it is. A passive carrier puts each drive straight under a root port; a switch card puts bridges of its own between the slot and the drives.
    lspci -tv
    ReadingEach drive hanging off its own root-complex port: a passive card, which depends on the slot split. Drives grouped behind one extra bridge that fans out to several ports: a switch card. The two switch-card makers we read state their cards do not need the split; for others, check the maker's requirements before ruling the split out. Root, upstream and downstream ports all report the same class, 06:04 'PCI bridge', so confirm each bridge's role with the next step. We found no source stating how a retimer card shows in this tree, so this page does not use the tree to tell a retimer card from a passive one.
  3. 3.Read the port type and link figures for each NVMe endpoint and for the bridge above it. Run it as root: on many systems part of PCI configuration space is readable only by root, and the capability lines can be missing otherwise.
    sudo lspci -vv -s <bus:dev.fn> | grep -E 'Express|LnkCap|LnkSta'
    ReadingThe 'Express' line names the role: 'Root Port (Slot...)', 'Upstream Port', 'Downstream Port (Slot...)' or 'Endpoint'. An Upstream Port with Downstream Ports beneath it is a switch. On each drive, LnkCap is its maximum (for example Width x4) and LnkSta what it negotiated; '(downgraded)' after the width means it trained on fewer lanes than it supports. Root ports and downstream ports never print '(downgraded)', so on the port above a lone visible drive compare the LnkCap and LnkSta widths yourself. We found no source stating that a root port's width reflects the slot's bifurcation setting, so treat a mismatch there as a lead, not proof.
  4. 4.Look for the kernel's bandwidth warning. When a device gets less PCIe bandwidth than it is capable of, the kernel logs the upstream link that limits it.
    sudo dmesg | grep -i 'PCIe bandwidth'
    ReadingA line '... available PCIe bandwidth, limited by <speed> x<width> link at <address> (capable of ... with ... x<width> link)' names both the speed and the width of the limiting link. A width below x4 on a drive's own link, or a narrow upstream link on a switch card, points at the path; a limit on speed alone points at the PCIe generation. It does not detect drives missing from a passive card in an x8-wired slot: each drive that came up there has its own x4 link, and a drive that never enumerated logs nothing.
Split the slot that holds the card
  1. 1.Find the slot's bifurcation option in the host firmware. It belongs to the board, not the card: Supermicro's FAQ for its AOC-SLG3-2M2 sends the reader to the motherboard manual to check and change it, and the card's own product page does not mention it. The only menu path we read in a vendor document is Lenovo's ThinkSystem UEFI for Xeon 6. The change is a firmware setting, not a change to the drives' data, but the machine reboots to apply it.
    System Configuration and Boot Management > System Settings > Devices and I/O Ports > Override Slot Bifurcation > Slot n (Lenovo ThinkSystem, Xeon 6)
    ReadingLenovo offers 'Auto PCIe Slot Bifurcation' (Disabled by default, or Enabled) and, per physical x16 slot, x16 (the default), x8x8, x8x4x4, x4x4x8 and x4x4x4x4. On Lenovo, x16 means the system setting is used; choose x4x4x4x4 explicitly. Lenovo notes that the values on offer vary with the hardware installed, such as the riser card.
  2. 2.Choose the value that gives every drive its own x4 link: x4x4 for a two-drive card, x4x4x4x4 for four drives in an x16 slot.
    ReadingSupermicro's statement for its passive two-M.2 card: a slot left at x8 or x16 shows one drive, x2x2x2x2 shows neither, x4x4 shows both. If the slot offers no value with one x4 per drive, it cannot run every drive on a passive card: try another slot (p3), and if no slot can, see the last cause below.
  3. 3.On other boards the option has other names, and the ones here come from user reports, not vendor manuals. On a Supermicro X12SPL-LN4F, ServeTheHome posters set it per IOU, and a reply says the drives behind the card show only with x4x4 set on the IOU that feeds that slot. Level1Techs replies say ASUS labels it 'PCIe RAID Mode' on the Pro WS X570-ACE. On an ASUS KRPA-U16 EPYC board, on BIOS 0302, which an engineering-sample CPU tied the poster to, the poster found no bifurcation setting at all, and an early reply suggested the board's OCuLink ports instead. Later posts share modified BIOS images that hard-set slot 1 from 'GLink1 x16' to x4x4x4x4, and mention a 'CPM Dxio Topology' menu item in a newer build; we did not evaluate either. We have not read a Supermicro or ASUS manual stating these names, and we found no Dell document stating its slot bifurcation option or values.
  4. 4.If installing the card upsets other onboard devices, set the split for its slot by hand rather than relying on the firmware's automatic choice. On a Supermicro X10SRH-CF, moving an NVMe card into slot 7 left both onboard NICs at 10/100 with no address. A reply placed slot 7 on the same IIO (#3) as the NICs, suspected the automatic bifurcation and suggested forcing a split. Another reply called it a BIOS bug. The poster reported 'It was bifurcation' and pointed to Advanced > Chipset config > North Bridge > IIo config > IIO 1 config, without saying which value fixed it.
  5. 5.If the value you expected brings the drives up only one at a time, try the other split values the firmware offers. On a Gigabyte MC12-LE0 with a Ryzen 7 PRO 5750GE, one poster saw one drive or the other at 4x4x4x4 and at 4x4x8, and both at 8x4x4. The thread does not explain why, and neither does this page.
Split is set, drives still missing: wiring, shared lanes, chipset, CPU
  1. 1.Read the slot table in the board's specification or manual and find the electrical width beside the connector size. Supermicro's X12SPL-F, for example, lists '1 PCIe 4.0 x16' beside '1 PCIe 4.0 x8 (in x16 slot)'. A four-drive passive card in an x16 connector wired x8 has eight lanes to divide, which by arithmetic is two x4 links, not four.
    Board specification or manual: the slot table and block diagram for the slot in use
    ReadingNo source we read says which of the four drives come up in that case, so do not expect a particular pair. The dmesg bandwidth line in p1 does not flag this case, so the slot table is the check.
  2. 2.Check what the slot shares. A slot's width can depend on what else is installed: Supermicro's X11SCA-F lists its two x16 slots as 'NA/16, or 8/8' and states 'M.2#1 is shared with PCIex4 slot, M.2#2 is shared with U.2'. ASUS's FAQ for its Hyper M.2 cards gives an example where four drives on PCIEX16_2 leave no lanes for PCIEX16_3. ASUS lists some slots that read only three drives and shows, as images, where to install them; the per-board list is an image too, so read both there for your board.
    ReadingEmpty an M.2 socket that shares the slot's lanes, or move the card to a slot that does not share them.
  3. 3.Find out whether the slot is fed by the CPU or by the chipset: the board's manual or block diagram says, and the two Supermicro spec pages above do not. Intel lists the Z790 chipset's PCI Express configurations as x1, x2 and x4, so a slot fed from that chipset is not an x8 or x16 link to divide.
    ReadingA chipset slot is not always a dead end: on an ASUS Pro WS X570-ACE a user split the chipset-attached third slot x4x4 for two drives. On an ASRock B650 PG Riptide, bifurcation did not work in the lower slot, and replies say the lower slots on that board are chipset-attached. Check your own board rather than a rule of thumb.
  4. 4.Check the splits the CPU itself can make. Intel's specification page for each processor carries a 'PCI Express Configurations' field: for the Core i9-14900K it reads 'Up to 1x16+4, 2x8+4' with 20 lanes, and no x4x4x4x4 configuration is listed. We have not read an AMD document listing its processors' configurations, so on an AMD platform go by the board maker's manual.
    ReadingLimits can also be a number of slots. In a ServeTheHome thread on a Dell Precision T7920, only one drive on a bifurcation card was recognised and only in two of the slots; a reply reports that only two x16 slots could be split at a time, with a third giving only its first M.2 drive, and one reply reports its best results with Intel VMD disabled and CMOS cleared after adding each card. No fix was confirmed for the original poster.
Every drive is on the bus, but the OS or a VM does not show it
  1. 1.If the count in p1 matches the drives on the card, the slot and the card have done their part. For a drive that is on the PCI bus with no /dev/nvme device, work through the driver and vfio-pci binding checks on our page for an undetected U.2 drive, linked below.
  2. 2.A host that passes the drives through to a VM can lose them for a reason that has nothing to do with the split. In a TrueNAS forum thread, four drives on an ASUS Hyper M.2 card in a Proxmox host, with the slot set to 4X4 and NVMe RAID off, went from three of four visible in a TrueNAS VM to one, then none. The journal showed vfio-pci 'Unable to change power state from D3cold to D0', and adding disable_idle_d3=1 to the vfio-pci options fixed it for that poster.

Causes, how to confirm each, and whether hardware fixes it

CauseHow to confirmFixHardware?
A passive card in a slot not split into one x4 link per drivelspci counts fewer class 0108 controllers than drives on the card; lspci -tv shows the visible drive directly under a root port; the slot's firmware setting reads x16 or x8. On Lenovo, x16 means the system setting is used; choose x4x4x4x4 explicitly.Set the slot to x4x4 (two drives) or x4x4x4x4 (four drives in an x16 slot) and reboot. If the drives then come up one at a time, try the other values the firmware offers.No
The slot is wired narrower than its connectorThe board's slot table lists the slot as x8 or x4 'in x16 slot', and lspci counts fewer drives than the card holds with the split set.Move the card to a slot whose electrical width covers one x4 link per drive.No
The slot's lanes are shared with an M.2 socket or another slotThe board spec lists shared lanes or widths that depend on what is installed ('NA/16, or 8/8', 'M.2#1 is shared with PCIex4 slot'), or ASUS's FAQ shows the slot pair for your board.Empty the shared M.2 socket, move the other card, or move the multi-drive card to a slot that does not share its lanes. On ASUS Hyper M.2 cards, check the slots and install positions ASUS's FAQ shows for your board.No
The slot hangs off the chipsetThe board manual or block diagram puts the slot on the chipset, and the chipset's configuration list has no x8 or x16 link (Z790: x1, x2, x4).Use a CPU-fed slot. Some boards can split a chipset slot, as one user reports for an ASUS Pro WS X570-ACE, so check your board's own options first.No
The CPU cannot make the split, or the platform splits only some slotsThe CPU's 'PCI Express Configurations' list lacks the split you need (i9-14900K: 'Up to 1x16+4, 2x8+4'); the firmware offers the split on some slots and not others.Use a slot and a split that the CPU and the board both list. Where only a set number of slots can be split, decide which cards get them.No
Intel VMD is enabled and the drives sit behind itThe count in p1 is short on an Intel platform with VMD enabled; the VMD check on our U.2 detection page shows whether the kernel has taken the VMD domain.Follow the VMD steps on the U.2 detection page. One reply in a ServeTheHome thread on a Dell Precision T7920 reports its best results with Intel VMD disabled; no fix was confirmed for the original poster.No
Drives passed through to a VM are stuck in a power stateThe drives were visible and then vanish from the VM; the host journal shows vfio-pci 'Unable to change power state from D3cold to D0'.One poster's fix was disable_idle_d3=1 in the vfio-pci options. For other binding problems, see the U.2 detection page.No
No slot on the board can be split for the cardNo slot with enough lanes offers a value with one x4 per drive, or the firmware has no bifurcation option at all, as one poster reports for an ASUS KRPA-U16 on BIOS 0302.Use a carrier with an on-board PCIe switch, which divides the lanes itself. HighPoint states its Rocket 1508 has no bifurcation dependency; Supermicro states its switch-based AOC-SHG3-4M2P fits a standard x8 slot with minimal, if any, BIOS or OS changes, and lists qualified platforms, all of them Supermicro systems and boards. Check the card maker's own host requirements.Yes
If no slot can be split: cards whose title says Bifurcation Built-in

A card that divides the lanes itself does not need the slot split: HighPoint and Supermicro each state this for their own switch cards. The block lists only multi-M.2 cards whose listing title says 'Bifurcation Built-in': that is the seller's wording, which we have not verified, and no title names a switch chip. On 2026-10-14 three products made that claim. Read each listing's own words, and match the host link it states to your slot's electrical width.

When not to buy: Do not buy when lspci already shows every drive, when the board has a slot that can be set to x4x4 or x4x4x4x4, or when the drives vanish only inside a VM: the fix is a firmware setting, a different slot or a binding. A passive card would need the same split and show the same single drive, which is why the block lists only cards whose title claims otherwise.

Every row below meets: Listing title states two or four M.2 drives on the card - model text contains "Dual M.2" · Listing title states two or four M.2 drives on the card - model text contains "Quad M.2" · Listing title states two or four M.2 drives on the card - model text contains "4-Drive M.2" · Listing title states two or four M.2 drives on the card - model text contains "Supports 4 NVMe" · Listing title states two or four M.2 drives on the card - model text contains "NVMe (4) M.2" · Title says Bifurcation Built-in (the seller's wording for a card that splits the lanes itself; not verified by us). A listing whose field is not stated is excluded from a required criterion and marked “not stated” on an optional one; nothing is assumed. What the catalogue cannot evaluate: whether the card needs the slot split; onboard PCIe switch; number of M.2 slots; card PCIe generation. Check these on the maker's data sheet.
Product · part numberCap.Interface · form · ratingConditionPrice$/TBBuy
GLOTRENDS PA20 Dual M.2 NVMe to PCIe 3.0 X4 Adapter
GLOTRENDS PA20 Dual M.2 NVMe to PCIe 3.0 X4 Adapter, Bifurcation Built-in
—n/a · AccessoryNEW Amazon: New / Unknown · sold by GLOTRENDS US Official Store$82.99—Buy on Amazon
GLOTRENDS ST560 Quad M.2 NVMe to PCIe 3.0 X8 Adapter
GLOTRENDS ST560 Quad M.2 NVMe to PCIe 3.0 X8 Adapter, Bifurcation Built-in
—n/a · AccessoryREFURB / USED Amazon: Used / LikeNew · sold by GLOTRENDS US Official Store · seller's note on the listing$161.10—Buy on Amazon
GLOTRENDS PA44 Quad M.2 NVMe to PCIe 4.0 X8 Adapter
GLOTRENDS PA44 Quad M.2 NVMe to PCIe 4.0 X8 Adapter, Bifurcation Built-in
—n/a · AccessoryREFURB / USED Amazon: Used / LikeNew · sold by GLOTRENDS US Official Store · seller's note on the listing$206.99—Buy on Amazon

Prices are Amazon US, re-checked about once a day; newest price in this table . A click here is a click, not a purchase; we earn a commission on qualifying purchases. The table shows what the catalogue holds against the criteria, not an endorsement of any model for your server.

Read next on this site

Frequently asked questions

Does a card with an on-board switch need the slot split?

Not according to the two makers we read. HighPoint states its Rocket 1508 has 'no-Bifurcation dependency or need for specific brand/model of motherboards' and can go in a PCIe 3.0 or 4.0 slot, bifurcated or standard. Supermicro states its AOC-SHG3-4M2P, which carries an on-board switch, fits a standard PCI-E 3.0 x8 slot with 'minimal, if any, BIOS or OS modifications', and lists qualified platforms, all of them Supermicro systems and boards. Each statement covers that maker's own card. Neither page names the switch chip, and this page does not either.

Which drives come up if a four-drive card sits in an x16 connector wired x8?

Eight lanes make two x4 links, so at most two drives can each have their own x4 link; that is arithmetic from the lane counts, not a vendor statement. No source we read says which positions on the card come up, so this page does not predict it. Move the card to a slot the board lists as x16 electrically.

Why does the same card work in one machine and not in another?

Because the usable split belongs to the board and the CPU, not the card. ASUS writes that 'Maximum M.2 SSDs support will vary depending on different CPUs and motherboards.' One ServeTheHome poster ran the same four drives split on a Dell Precision T7810 and could not get them all on a T7920.

Do I need to buy anything?

Usually not. A slot split, a different slot, an emptied M.2 socket, a VMD setting or a vfio-pci option fixes most of the causes above. A new card helps only when no slot on the board can be split, and then only a card that divides the lanes itself. HighPoint and Supermicro state that in their own documents for their own switch cards; the cards listed on this page say only 'Bifurcation Built-in' in their listing titles, which we have not verified.

Sources

Every technical statement above rests on one of these documents, each read on the date shown. Where a document did not state something, the page says so rather than filling the gap.

  1. Supermicro FAQ 25062 - AOC-SLG3-2M2 in a PCIe slot that cannot be bifurcated · checked
  2. Lenovo ThinkSystem UEFI (Xeon 6) - Override Slot Bifurcation · checked
  3. HighPoint Rocket 1508 8x M.2 Port to PCIe 4.0 x16 NVMe Switch AIC - product page · checked
  4. Supermicro AOC-SHG3-4M2P add-on card product page · checked
  5. Supermicro AOC-SLG3-2M2 add-on card product page · checked
  6. Supermicro X12SPL-F motherboard specifications · checked
  7. Supermicro X11SCA-F motherboard specifications · checked
  8. ASUS FAQ 1037507 - Compatibility of PCIE bifurcation between Hyper M.2 series Cards and Add-On Graphic Cards (last update 2026/02/05) · checked
  9. Intel Core i9-14900K - ARK specifications · checked
  10. Intel Z790 Chipset - ARK specifications · checked
  11. lspci(8) man page · checked
  12. pciutils ls-caps.c - cap_express(), LnkCap/LnkSta printing and link_compare() · checked
  13. Linux kernel documentation - The PCI Express Port Bus Driver Guide HOWTO · checked
  14. PCI ID Repository - Class 01 subclass 08 and class 06 subclass 04 · checked
  15. Linux kernel drivers/pci/pci.c - __pcie_print_link_status() · checked

Where people hit this

Discussions we opened and read while researching this page. They are evidence that the problem occurs and of what people tried; no specification above is sourced to a forum post.

Other server compatibility and memory problems

Share:WhatsAppFacebookPost on X