BIOS: ASUS EX-B560M-V5 (desired state)

The BIOS settings of this board are not declarable from NixOS: they live in the firmware NVRAM, and the kernel firmware-attributes interface (used by fwupdmgr get-bios-settings) only exists on Dell/Lenovo/HP business boards, not on consumer ASUS. fwupd here answers literally This system doesn't support firmware settings.

So this file is the closest thing to "declarative" available: the intent, versioned. It does not apply itself, but it survives a Clear CMOS and it is reproducible by hand in ~2 min.

Machine: Intel i5-11400 + Arc B580 (Battlemage). BIOS 2803 (2025/12/26).

SettingsCopy Anchor Link

Access: Del at POST, then F7 (Advanced Mode).

• Confirmed (OS): measured from Linux. CSM off (the Arc is UEFI-only and it booted), Above 4G + ReBAR active (a 16G VRAM BAR mapped above 4G, through lspci).
• Recommended: good practice for the Arc; not verified, confirm on the next visit.
• XMP: the TeamGroup UD4-3200 kit was running at 2400 MT/s (XMP off); with XMP on it is confirmed at 3200 MT/s (dmidecode -t 17).

Why CSM = Disabled is criticalCopy Anchor Link

Arc GPUs have no legacy VBIOS (they are UEFI-only). With CSM on, the motherboard tries to bring video up in legacy mode, finds no ROM on the card and hangs on the ASUS logo. That was the "CSM Forced Enablement Phenomenon" documented on the Intel forum, and it is not a defect of the board.

The root of the problem on this board: the EX-B560M-V5 had a default that re-enabled CSM automatically (CSM came back by itself after a reboot or a failed POST). That default was turned off, which is the real fix, at the source. With CSM off and that auto-enable off, you can enter the BIOS normally with the Arc installed (video comes up through the card's GOP, with no broken screen). The reboot=pci (OS side) stays as reinforcement.

TODO: write down here the exact name of the auto-CSM option that was turned off (its BIOS label), for precise reproduction after a Clear CMOS.

The OS side (this part IS declarative)Copy Anchor Link

• boot.kernelParams = [ "reboot=pci" ] in modules/nixos/hardware/intel-arc-b580.nix: it forces a full reset through 0xCF9 on reboot, the GPU reinitializes clean as in a cold boot and the POST does not fail, so CSM is never turned back on. It fixes the warm-reboot hang.
• services.fwupd.enable = true in modules/nixos/hardware/hardware.nix: it does NOT cover this board's BIOS (ASUS is outside LVFS); it serves SSD firmware and other components.

The Arc's idle power: ASPM on the CPU link (18/10/2026)Copy Anchor Link

The B580 drew 45 W on the board with the desktop idle (its own xe sensor, read by the NixOS Center). The usual cause on Arc is a link that never enters ASPM L1, and lspci -vv (as root) showed exactly where it stops:

The card is ready; it is the CPU's PEG port that ADVERTISES no ASPM, and Linux honors the advertisement. So the fix is the firmware's, NOT a kernel parameter: pcie_aspm=force would enable a state the port says it does not have, which is how links hang.

The firmware has the options and hides them (VERIFIED 18/10/2026 against BIOS 2803, the official EX-B560M-V5-SI-2803.zip from dlcdnets.asus.com/pub/ASUS/mb/BIOS/, unpacked with uefiextract and its Setup module (899407D7-...) read with ifrextractor, nothing written). The options exist in its forms: PCI Express Native Power Management and Native ASPM (Setup), PEG - ASPM with Disabled/Auto/L0s/L1/L0sL1 (SaSetup), the PCH ASPM (PchSetup). But in the Advanced form, the links to Platform Misc Configuration, System Agent (SA) Configuration, PCH Configuration, Power & Performance, RC ACPI Settings and Thermal Configuration all sit under SuppressIf 1 == 1, a condition that is always true. What stays visible is exactly the manual's list (CPU, PCH Storage, PCH-FW, PCI Subsystem, USB, Network Stack, NVMe, Onboard Devices, APM, SMART), so an Expedition board simply does not offer these knobs. The SA menu being hidden also means the Primary Display row above cannot be set from the menus either.

The live values agree (read from efivarfs, 4-byte attribute header first; Setup is 2698 bytes = the IFR's 0xA86 plus 4, so the offsets match this BIOS):

There are TWO PEG - ASPM questions. The form picks one by the CPU: SaSetup 0x52 is suppressed when CpuSetupVolatileData reads 0x670 (CPUID A0671 masked, Rocket Lake), and SaSetup 0x35D, the "PCI Express Root Port 1" page, is the one shown for this CPU. L1 Substates (0x361) stay off: a Battlemage card has been reported failing to boot with L1.1 enabled.

The exception: bios-aspm (decided 18/10/2026, review 18/01/2027)Copy Anchor Link

The "Do NOT do" rule against raw variables gets ONE exception, these three bytes, by decision on 18/10/2026: the menus cannot reach them, the gain is the largest idle draw on the machine, and the risk is bounded by the checks below and a Clear CMOS. hosts/ex-b560m-v5/bios-aspm.nix installs bios-aspm:

• bios-aspm status (no root): the three values, whether the kernel still logs the FADT line, and as root the 00:01.0/01:00.0 link state.

• sudo bios-aspm apply: refuses on any other board, any BIOS but 2803, a variable of another size, attributes other than 07000000, or a current value outside the option's valid set. Then it backs both variables up to /var/backup/bios-aspm/<date>/ with SHA256SUMS (inside restic's paths), asks for yes, writes each variable in ONE write as efivarfs requires, reads it back, and puts the backup back on any mismatch.

• sudo bios-aspm restore <dir>: writes a verified backup back.

• The guard, a dead man's switch. apply (or sudo bios-aspm arm <dir>) writes /var/lib/bios-aspm/pending = this boot's boot_id plus the backup. The timer bios-aspm-guard runs 10 min into every boot and, on a boot OTHER than the arming one, restores that backup unless sudo bios-aspm confirm removed the file first. So a boot that comes up with no screen and no SSH heals with a 10 min wait and a power cycle. The boot_id matters: the arming boot's own timer has long elapsed and fires at once, and must not undo the change before the reboot. A failure of the guard raises a my.alert. TESTED 18/10/2026 on copies: same boot left it alone; a fake later boot_id restored the backup and cleared the file; confirm left the guard idle.

• The watch, while it runs (bios-aspm-watch, a service following journalctl -k -b -f, event-driven). It only judges the card's path: the root port 00:01.0, the bridges 01:00.0, 02:01.0, 02:02.0, the GPU 03:00.0 and its audio 04:00.0.

  • An uncorrectable AER error, a Link Down, or the xe driver declaring the device wedged: it writes the newest checksummed backup back at once and raises a my.alert. Nothing changes until the next boot, which then runs without ASPM. To try again, apply.
  • Corrected AER errors on that path: an alert at most every 10 min with the count. The hardware recovered, but a steady stream means L1 is marginal on this link.
  • An xe GT reset or a timed-out job: an alert with the restore command. A single one can be a game's fault; repeated ones are the signal to roll back.
  • At its start, if the bytes were seen applied before and are gone now (an F5, a Clear CMOS, a BIOS update), it says ASPM is off again; the marker is /var/lib/bios-aspm/enabled. TESTED 18/10/2026 on copies with fed kernel lines: the second corrected error within 10 min stayed quiet, an error on the NVMe's port 00:06.0 was ignored, a fatal one on 00:01.0 restored the backup and cleared the marker, and a lost-bytes start raised the alert.

It does NOT cover a board that fails to POST, since Linux never runs. Before that jumper: the logo screen's Del then F5 (Load Optimized Defaults) resets every setup variable, hidden ones included; and with no image from the B580, the board's own HDMI/DP (the i5-11400's iGPU) is where the firmware falls back.

TESTED before any write, 18/10/2026: the patch applied to copies of the live variables changed exactly positions 86 and 87 of Setup and 866 of SaSetup (0x51, 0x52 and 0x35D plus the 4-byte header), sizes intact. Two traps it already avoids: cp cannot read efivarfs (Illegal seek), so it copies with cat; and grep -q after journalctl under pipefail reads a match as a miss.

APPLIED 18/10/2026 20:49, rebooted 21:05, confirmed 21:11. After the reboot the FADT line is gone, _OSC hands the OS PCIeHotplug PME AER PCIeCapability LTR, the root port's LnkCap reads ASPM L1, Exit Latency L1 <16us (it read ASPM not supported), and both ends of the card's link read ASPM L1 Enabled. No AER error and no xe reset in that boot. The idle watts are PENDING: the readings after the reboot (55, 53 and 50 W on the board) were taken with Zen redrawing all the time, the GT at 0 to 1% RC6, so they measure a busy GPU, not an idle one. The fair reading is the GPU BOARD card with the screen quiet and RC6 high.

On the review date: if ASPM held with no xe resets or link errors, the exception stays and this paragraph gets the measured watts; if anything regressed, restore and drop the module. A BIOS update changes the offsets: bios-aspm then refuses, and the IFR has to be read again.

The kernel log names the same cause before any lspci. On the boot of 18/10/2026: ACPI FADT declares the system doesn't support PCIe ASPM, so disable it, FADT indicates ASPM is unsupported, using BIOS configuration, and r8169 ... can't disable ASPM; OS doesn't have ASPM control. That FADT flag is what Native ASPM = Disabled publishes, so its absence is the first sign the BIOS change took.

Checking it, after the BIOS change and a reboot (read only):

journalctl -k -b | grep -iE 'FADT.*ASPM|OS now controls'      # wanted: no FADT line
P=$(nix build --no-link --print-out-paths nixpkgs#pciutils.out)
sudo $P/bin/lspci -vv -s 00:01.0 | grep -E 'LnkCap|LnkCtl'   # wanted: ASPM L1 in both
sudo $P/bin/lspci -vv -s 01:00.0 | grep LnkCtl               # wanted: ASPM L1 Enabled

Then read the board's idle watts again (the NixOS Center's GPU BOARD card, desktop idle, both monitors on), and watch a day of normal use and one game session for stutter, a black screen or an xe reset in journalctl -k. If anything regresses, set PEG - ASPM back to Disabled: that is the whole rollback. Two lit monitors at 180 Hz and 144 Hz keep the display engine busy, so even a working L1 will not bring the card down to single-monitor figures.

Backup and restoreCopy Anchor Link

• Profile in the BIOS (native): Tool → ASUS User Profile (or Save & Exit → Save Profile), which saves into a slot (1-8) in NVRAM. After a Clear CMOS it is just Load Profile.
• Recovering from a hang: hold the power button ~8s, and a cold boot always recovers the Arc. As a last resort, a Clear CMOS (the CLRTC jumper, or pulling the battery for ~1 min) goes back to defaults.

Do NOT doCopy Anchor Link

• Do NOT flash a modified BIOS (a mod for hidden menus): this board has no USB BIOS FlashBack, so a brick becomes a job for an external SPI programmer (CH341A + clip).
• Do NOT edit the raw variables (except the three bytes of bios-aspm above) (AMITSESetup, CpuSetup and friends, through setup_var/efivarfs by offset): the offsets are undocumented and change per BIOS version, so the risk is corrupting or bricking it. The options that matter are already in F7 Advanced Mode.

Updating the BIOS (reference)Copy Anchor Link

1. Download the version's .CAP from https://www.asus.com/motherboards-components/motherboards/expedition/ex-b560m-v5/helpdesk_bios/ and check the official SHA-256.
2. A FAT32 flash drive, with the .CAP at the root.
3. Del → F7 → Tool → ASUS EZ Flash 3 Utility → pick the file.
4. The flash resets the settings to defaults, so reapply the table above afterwards.