Host firmware (MMIO High / large BAR / SR-IOV)
WIP / community: Workstation BIOS work for 4× V620 + PLX. Reports are from
#motherboardand#general(Sep 2026). Treat every offset as your dump, not a copy-paste from a PDF. See Verification status.
Large-BAR cards (V620 is 32 GB GDDR6, plus a huge prefetchable MMIO window) can fail to POST on older workstation firmware even when the same cards boot on a modern desktop board. This page is about getting the host to enumerate the GPUs — not about PCIe P2P after Linux is up.
When this applies
Community symptoms:
- Host POSTs fine with no GPUs (or one GPU) and hangs with 4× V620 behind a PLX / switch
- bifurcation card.
- The same 4× V620 + PLX kit works on a recent consumer board (community: Z690-class) but not on a Dell Precision tower you want for more RAM / more slots.
lspcinever appears because firmware never hands off.
If Linux already sees all cards, skip this page and go to P2P / Getting started.
Dell Precision T5820 / T7820 / T7920 (community)
#motherboard + #general (Sep 12–14 2026): these Xeon-W / dual-socket Precision towers often need
a hidden MMIO High / large-BAR change. A circulating public write-up is the
Precision 7920 MMIO GPU Unlock Guide
(GitHub attachment). Community: T7820 behaves like T5820 for this class of tweak.
Operational rules (do these)
- Dump your own UEFI variables. Do not copy a hex blob from the PDF or from someone else’s BIOS version. Community first-fail: writing another host’s dump back.
- Confirm the MMIO change without the PLX (cards in CPU slots, or fewer cards). Then add the switch.
- Prefer EFI
.nshscripts over typing hex by hand. Community usedsetup_var.efiand UEFITool against the matching BIOS update package (example: T7820 2.52.1) to find hidden variables. - Community variable name (T5820):
SocketCommonRcConfig, sub-fieldmmiohsizeset to0x03or0x04. Write base + size together; keep the existing base. Both sizes have been reported as “the one that worked” — A/B on your dump. - Do not factory-reset, clear CMOS, or pull the CMOS battery after a successful write — that
clears the change.
#general(Sep 23): a host that lost BIOS settings (flat CMOS battery suspected) came back with SR-IOV re-enabled and ReBAR off, then saw MMIO not allocated / fewer GPUs than before. Re-apply the known-good firmware settings before chasing ROCm. - After Linux boots, continue with P2P readiness.
Aperture size (community)
V620 VT-d / IOMMU windows reserve a very large prefetchable range per card (community order of magnitude: hundreds of GB of address space per GPU, not “just 32 GB VRAM”). A 64 GB MMIO High window can be too small for 4× V620. Community: there is often no 128 GB option; jump to 256 GB if the menu / variable allows it. Mixed large-BAR cards (e.g. adding 32 GB MI50s next to V620s) made a too-small window worse until 256 GB was set.
Dual-socket towers
T7820-class boards can route some slots to CPU2 or the chipset. Label the slots and keep all inference GPUs on one CPU’s root ports when you can. See Host topology.
What actually unblocked 4× V620 + PLX (one host)
Community (#motherboard, Sep 14 2026): Dell Precision T7820, 4× V620, LSI PEX880xx-class
switch on a bifurcation card, guest Proxmox.
- MMIO High /
mmiohsizework was necessary but not sufficient — POST still failed with the PLX populated. - The change that then posted and booted was disabling SR-IOV via hidden UEFI variables located with UEFITool on that BIOS package.
- That SR-IOV fix is board-specific. Do not treat it as a recipe for ASUS / HP / consumer boards.
Other platforms (short)
| Platform | Community note |
|---|---|
| Supermicro H12D | #general (Oct 6 2026): one H12D + V620 host looped POST codes 95, DE, AD, DF, 36, 05, 37 (reporter thought the card was dead). The combo that then POSTed: SR-IOV off, Above 4G decode on, CSM off, ReBAR on. Same thread: SR-IOV off was also needed for 4 GPUs per PLX on that board. Needs verify on other H12D BIOS revs — see below. |
| HP Z4 G4 | Firmware described as 32-bit MMIO with V620 — BAR / crosstalk pain; dual-CPU Z6/Z8 G4 are a different (more complex) story. #motherboard (Sep 16), Xeon W-2133 host: BIOS PCIe MIMO Assignment = 32-bit and ReBAR on. That 32-bit ceiling would fail to map 64 GB of VRAM; Linux then reallocates with GRUB pci=realloc=off pci=nocrs iommu=pt amdgpu.aspm=0 amdgpu.gpu_recovery=1. Needs verify on other Z4 BIOS revs. |
| HP Z6 G4 | #motherboard (18–23 Oct), single-CPU / 96 GB host, Fedora or Ubuntu: after a BIOS upgrade, front / memory-header blower fans could sit at the wrong PWM until a BIOS factory reset (settings looked pegged at 100%). Same thread: a second V620 in slot 5 failed with PSP load kdb failed! / PSP init error; that V620 POSTed in slot 2, and a small Quadro in slot 5 was fine. MMIO 32-bit, ACS off, ReBAR, PCI reset, and the Z4 GRUB realloc line were already set. #motherboard (29 Oct): community slot map slots 2 and 5 = x16, slot 4 = x8; a hidden BIOS setting (not in the visible menu) is the current guess for why a second V620 fails to init. Same thread (26 Oct): another Z-class host put V620s in slots 1 and 3 (x16) and a Quadro in slot 5 (x8). Unresolved / Needs verify — not a dual-V620 recipe yet. |
| ASUS WS-Z270 + ReBAR | #motherboard WIP on MI50 (not V620). Above 4G on, CSM off, ReBARUEFI. One card at 32 GB ReBAR, two cards collapsing to 16 GB / POST code 96 is not a gfx1030 recipe. Sep 15: ReBarState > 14 with two 32 GB cards threw code 96; UEFIPatch hit PciBus only, not PciRootBridge. Suggested A/B: GpuMMIOFix (OS remap, typically every boot) — Needs verify. |
| Modern desktop (Z690-class, etc.) | Often POSTs 4× V620 + PLX without hidden MMIO patches (community contrast to Precision). |
Supermicro H12D
#general (Oct 6 2026): dual-SP3 H12D (community also uses H12D-8D as a P2P / bifurcation host) can fail to POST a V620 with a repeating debug-LED loop. Community first-success set:
- SR-IOV — disabled (also reported required for 4 GPUs behind one PLX on this board).
- Above 4G decode — on.
- CSM — off.
- ReBAR — on.
That is a menu-level combo, not the Dell hidden-mmiohsize path above. Do not factory-reset /
clear CMOS after it works — on other boards that re-enables SR-IOV and turns ReBAR off. Needs
verify on other H12D revisions.
BIOS editors and unsigned flashes void warranty and can brick the board. Use the vendor flashback / recovery path if you have one. This wiki does not publish byte patches.
Tools (public)
| Tool | Use |
|---|---|
| UEFITool | Decode the same BIOS update you are running; find hidden setup variables. |
| ReBARUEFI | Expose / patch ReBAR + related MMIO options when the menu hides them. |
| GpuMMIOFix | OS-side BAR remap above 4 GB (#motherboard Sep 15–16). Community: load each boot. Needs verify — not a substitute for a correct MMIO High / SR-IOV firmware fix on Precision towers. |
| Vendor BIOS .exe / capsule for your revision | Source of truth for variable layouts. |
Related
- Supported hardware — card identity
- Multi-GPU P2P — after the host POSTs
- Getting started
- Useful resources