Operating systems and kernels

Reported 8 GB and 10 GB

Which operating systems, kernels and driver versions people run unlocked CMP 170HX cards on, and what broke. Linux with the patched open kernel modules is the normal path. The unlock itself is covered on Unlocks; this page is about the host software around it.

The short version

Any mainstream Linux works if it can build the patched NVIDIA open kernel modules: Ubuntu 22.04 and 24.04, Proxmox VE 9.2 (Debian based), CachyOS and EndeavourOS (Arch based) all appear below.
Pin the kernel and the NVIDIA packages once it works. A kernel or driver package upgrade puts stock modules back through DKMS and the cards silently drop to 8 GB and Gen1.
Secure Boot off. The patched module is not signed by a key your firmware trusts.
Driver: nvidia-open 615.71.09 is the version most current builds target.
Windows works on the 10 GB card with the A100 40G / A800 40G driver; no 8 GB Windows unlock is reported.

Reported working

OS	Kernel	Driver	Unlock build	Setup	Source
Ubuntu 24.04	6.8.0 (6.8.0-147-generic)	nvidia-open 615.71.09	Morrowmake/cmpunlocker `p2p-ecc` `87cb5fd`	bare metal, 8 cards, Gen2 x16, 64 GB BAR1, P2P	Source: @darkmaniac7, 2026-10-29 (lab notes)
Ubuntu 24.04 (guest)	not recorded	nvidia-open 580.178.04 with the v0.5 core ported	own port	ESXi 8.0.3 passthrough, 4 cards	Source: @darkmaniac7, 2026-10-21 (lab notes)
Proxmox VE 9.2 (host) + Ubuntu 24.04.4 (guest)	host 7.0.x-pve, guest 6.8.0-139-generic	host 610.43.02 patched, guest stock 610.43.02	amoghmunikote/cmpunlocker `76f0954`	VFIO passthrough, unlock on the host	Source: repo github.com/anoane/01-proxmox-dual-gpu-passthrough by anoane on GitHub, posted by @logy0, 2026-10-10
Proxmox VE 9.2 (bare metal, GPU LXC)	7.0.2-6-pve	nvidia-open 615.71.09	cmpunlocker `88e39ce` + `6c442ee` + club patches	4 cards behind two PLX PEX 8747 switches	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @djpandaman, 2026-10-24
Ubuntu Server 22.04 (guest)	6.8	610.43.03 open	cmpunlocker `7019bc2`	Proxmox VFIO, Q35 + SeaBIOS	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @djpandaman, 2026-10-24
Proxmox (Debian based)	not given	not given	not given	"it works on my proxmox system"	Source: @galygious, 2026-10-10
Debian 13.4 (Trixie), VM	6.12.113+deb13-amd64	nvidia-open 615.71.09	cmpunlocker `2d10f89`	PCI passthrough guest; build fix and the compute-hang fix below	Source: @vipic, 2026-10-18
CachyOS (Arch)	not given	not given	not given	desktop with an RTX 4080 Super beside the CMP	Source: @ybryany, 2026-10-14 Source: @ybryany, 2026-10-14
EndeavourOS (Arch)	LTS kernel	oldest driver cmpunlocker supports	not given	one 10 GB card, used for the RIR tool	Source: @lycorascence, 2026-10-23
Windows	n/a	A100 40G / A800 40G driver	not given	10 GB card only	Source: @wakeupmbrc1, 2026-10-11 Source: @arabus_74064, 2026-10-16

The headless pattern is common: mrpandagod runs omarchy (Arch + Hyprland) on his desktop and only SSHes into the headless CMP machine; he notes omarchy's non-standard config files make troubleshooting harder. Source: @djpandaman, 2026-10-10
Source: @djpandaman, 2026-10-10

Reported failing or needing a fix

Kernel 7.2.11: Lycorascence (10 GB card, after an RIR burn) had cmpunlocker's link retraining break until he moved off 7.2.11 and rolled back the NVIDIA driver. He suspects a clash with Zagitta's gpu_unlock module, "possibly not aiming just for 3090 cards". Source: @lycorascence, 2026-10-23
Debian 13: on Debian 13.4 (kernel 6.12.113+deb13-amd64, nvidia-open 615.71.09, cmpunlocker 2d10f89) the module build failed with fatal error: stdarg.h: No such file or directory. ViPiC fixed it in driver/build.sh by pointing the kernel source and build trees at separate paths (KSRC=/lib/modules/${KVER}/source, KSOUT=/lib/modules/${KVER}/build) and passing both to make as SYSSRC and SYSOUT. The Debian 13 CUDA apt repository offered 615.71.09, 610.57.04 and 610.43.02 but not 615.78.08 (2026-10-18). Summary of his write-up; see Discord. Source: @vipic, 2026-10-18
Source: @vipic, 2026-10-18 clee maintains a fork of asm64-hooligan's cmpunlocker with a Debian build fix. Source: @c133, 2026-10-22 dagb, coming from Gentoo, found installing the NVIDIA drivers on Debian much harder than expected. Source: @da6b, 2026-10-23
Ubuntu 26.04, kernel 7.0.0-38-generic, stock 610.57.04: one new 10 GB card never initialized (kgspExtractVbiosFromRom_TU102: did not find valid ROM signature, VBIOS read as 0xffff). That was a card fault (bad SPI flash suspected), not the OS: it failed on the stock driver too. Source: @youngju10000, 2026-10-08
Source: @ichundes, 2026-10-08 See Card not detected or BAR allocation fails.
RIR tool access: the 170hx-rir-tool needs the kernel parameter iomem=relaxed to read anything, the cmpunlocker-patched driver (or a 610+ driver), and on some kernels the LTS kernel, where CONFIG_IO_DEVMEM_STRICT does not block it. Source: @lycorascence, 2026-10-20
Source: @lycorascence, 2026-10-20
Source: @lycorascence, 2026-10-23
Driver 615.78.08: the cmpunlocker v0.5 ECC build failed on 615.78.08 for Kaatchi and HenryKrinkle; 615.71.09 worked. Proven Source: @catchyuser, 2026-10-20
Source: @.henrykrinkle, 2026-10-20
615.71.09 with other GPUs in the same guest: under ESXi, other NVIDIA GPUs in the same VM hit krcWatchdog: GPU is probably locked and UVM wedged; NVIDIA's 615.71.09 notes list that symptom as a known issue. The fix there was porting the unlock to 580.178.04. Source: @darkmaniac7, 2026-10-21 (lab notes)

Driver versions

nvidia-open 615.71.09 was added in cmpunlocker v0.4; v0.5 was current on 2026-10-06. Source: @xeax, 2026-10-06
Source: @xeax, 2026-10-06
Source: @xeax, 2026-10-06
610.43.02, 610.43.03 and 610.57.04 are the earlier supported versions (cmpunlocker 76f0954 lists 610.57.04, 610.43.03 and 610.43.02). Source: repo github.com/anoane/01-proxmox-dual-gpu-passthrough by anoane on GitHub, posted by @logy0, 2026-10-10
615.78.08: see the ECC build failure above.
Use the driver version your unlock build lists, not the newest one your distro offers. See Unlocks.

Pinning, DKMS and Secure Boot

A kernel or nvidia package upgrade puts the stock modules back through DKMS. The cards drop to 8 GB and Gen1 without any error. Hold the kernel and every nvidia package once it works, and re-run the unlock install after any kernel upgrade. Source: @darkmaniac7, 2026-10-29 (lab notes)
Under ESXi, identify modules by sha256, not srcversion. Source: @darkmaniac7, 2026-10-21 (lab notes)
A stock DKMS copy under updates/dkms can shadow the patched module under updates/cmpunlocker; the card then shows 8 GiB. Fix: dkms remove nvidia/<version> --all, depmod -a, update-initramfs -u, check that modinfo -n nvidia points at the cmpunlocker module, then power off fully. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @djpandaman, 2026-10-24
Secure Boot must be disabled: mokutil --sb-state should report disabled; the patched module is self-signed at best. Source: repo github.com/anoane/01-proxmox-dual-gpu-passthrough by anoane on GitHub, posted by @logy0, 2026-10-10
nvidia: module verification failed: signature and/or required key missing - tainting kernel in dmesg with the patched module: @matstatman called it harmless ("its alright"). @xanthefin first did not see the line in his own dmesg, then found it there too. Source: @xanthefin, 2026-10-18
Source: @matstatman, 2026-10-18
Source: @xanthefin, 2026-10-18
A "no ECC support" line in dmesg from the ie31200 EDAC driver is about the host memory controller, not the GPU. Source: @mowgli8628, 2026-10-18
Source: @matstatman, 2026-10-18

Kernel command line

Bare-metal reference box: amd_iommu=on iommu=pt, nothing PCI-specific. Source: @darkmaniac7, 2026-10-29 (lab notes)
Proxmox host: quiet amd_iommu=on iommu=pt. Source: repo github.com/anoane/01-proxmox-dual-gpu-passthrough by anoane on GitHub, posted by @logy0, 2026-10-10
pci=realloc only when the BIOS sizes bridge windows too tightly for the 64 GB BAR1; see Server BIOS setup case C.
iomem=relaxed only for tools that read card registers through /dev/mem (the RIR tool above).

Mixed GPUs in one host

The unlocked driver drives every NVIDIA GPU in the host, so other cards must be supported by the same driver branch: Turing (RTX 20-series) or newer. Source: @exile_1ck, 2026-10-13
Source: @lenxtroz, 2026-10-13
Pascal display cards (GT 1030, GT 710) get no driver from these branches and fall back to 1024x768 basic output; run headless or use the board's BMC VGA. Source: @68x0, 2026-10-12
Source: @lia_oc, 2026-10-12
NVreg_RegistryDwords keys such as RmForceEnableGen2=1;RMPcieLinkSpeed=0x1 are global and also force other NVIDIA cards in the host to Gen2. 68x only tested them on a CMP-only rig; ybryany left them off on his RTX 4080 Super desktop. Source: @68x0, 2026-10-13
Source: @ybryany, 2026-10-14
Zagitta's consumer tensor unlock (gpu-unlock) and cmpunlocker should coexist: the gpu-unlock module is driver-version agnostic, runs before the NVIDIA driver loads and is gated by PCI device ids; not yet tested by the person who asked. Theory Source: @68x0, 2026-10-26
Source: @68x0, 2026-10-26
One model across different GPU types. Asked whether vLLM runs on a mixture of GPUs, Mountains and Con3po said yes; Con3po has done it across a 3090 ("3090 xtx" as written), an AMD 395 system and a Blackwell 4000. No run of one model across CMP and V100 cards has been posted: ElectroHobo had both in one Dell R730 under Proxmox but in separate VMs, and thinks the only way is two VMs plus a framework that shards the model across machines. CMP plus RTX PRO 6000 in one vLLM job is on Engines#Mixing CMP and newer GPUs in one job; VMs on Virtual machines and passthrough. Source: @segmond, 2026-10-25
Source: @mountains1605, 2026-10-25
Source: @con3po7016, 2026-10-25
Source: @nathang6866, 2026-10-25

Windows

The 10 GB card unlocks on Windows using the A100 40G / A800 40G driver. Source: @wakeupmbrc1, 2026-10-11 arabus posted a screenshot of it working on 2026-10-16. Source: @arabus_74064, 2026-10-16
No 8 GB Windows unlock is reported in these sources. Changing the 8 GB card's device id to A800 to use A800 drivers on Windows is an idea, not a result; see Unlocks.
Big Ptoughneigh reports game performance on the CMP 170HX is far higher on Windows than on Linux (about 60 vs 20 fps) and suspects the open driver. Source: @bigptoughneigh, 2026-10-16