Gen2 by platform
More actions
Gen2 doubles the PCIe link rate of a CMP 170HX from 2.5 to 5 GT/s, and whether it trains depends far more on the host than on the card. This page collects the reported results by board, CPU, operating system, kernel and driver, from the first Gen2 branch in July 2026 to early October 2026: a table first, then the detail behind each row. How the unlock works and how to install it is on Gen2 link retrain; how to check the link is on Verifying an unlock. Most rows date from July to September 2026 and name the branch or fork where the post did, so retest a board that failed on an early branch with a current release before you call it bad. Gen2 inside a virtual machine has its own section below and its own page, Proxmox Gen2 passthrough.
Before you compare
Gen2 is a link-speed unlock, independent of width; it needs a cold power cycle and does not train on every host
Gen2 raises the link rate from 2.5 to 5 GT/s and leaves the width alone: the same retrain reaches Gen2 at x4, x8 or x16. Width comes from the slot, bifurcation and the PCIe x16 capacitor mod, never from the speed patch ("an x4 connection cannot be made x16 by this service"). The patched driver opens a short Gen2 window while it brings the card up, a few seconds after it loads, and the link has to be retrained inside that window; the early-boot retrain service in the published unlocker exists to catch it. After installing, power the machine fully off for 60 s: a warm reboot does not reset the card's state. Judge the result by the link status in PCIe config space (lspci -vv LnkSta, or setpci -s <bdf> CAP_EXP+12.w: 0x1042 = Gen2 x4, 0x1041 = Gen1 x4), not by nvidia-smi, which can keep showing gen 1 after a successful retrain. Some Intel platforms never reach Gen2 on any branch or slot while AMD hosts reach it with no kernel command-line change; the discriminator is when the retrain fires relative to the window, not the IOMMU mode. rjohny55's unlocker retries a bounded number of times instead of looping: on a single Xeon E5 v4 host one card went from Gen1 x16 to Gen2 x16 only after a function-level reset and a second pass, and on an AMD EPYC host eight cards reached Gen2 x16 on the first pass after a reboot, all keeping 65536 MiB; a probe-serialization patch for driver 615 brought all eight back after an async-probe lock failure. On some board revisions the driver log shows FAILED to set OPT_GEN23: 64 GB and the SM unlock still work on those cards, but Gen2 cannot. Gen3 is fused off; forcing 8 GT/s on such a card drops it off the bus until the chassis is power cycled.
History: Gen2 first ran on Discord with Amogh's early gen2 branch in July 2026, which failed with two cards; Newman's multi-card retrain changes (AMD, HiveOS, two cards at Gen2 x4, 2026-07-24) were the fork others used. Who wrote the retrain patch (0008) is disputed: the lab's sweep credits the bendy2 fork, Eksitus called it "Newman's 0008 patch", and aifonder reached Gen2 with "bendy2's repo" (2026-07-29). oleh borzykh's capture matches the window: the watcher from cmpunlocker PR #35 saw the card change its advertised speed 6.262 s after load and retrained it from Gen1 x4 to Gen2 x4 (2026-09-01). dottoreargus noted that nvidia-smi shows the device maximum as Gen1 even while Gen2 is negotiated (2026-09-07). The first reports of Intel boards that never reach Gen2 are below: the ASUS X99-A and Z10PA-D8, the W890 SAGE and the Rampage VI Extreme Omega (July), and an X99 C612 board (August).
Proven 8 GB and 10 GB
Source: @newman_08462, 2026-07-24
Source: @gmzombie, 2026-07-24
Source: @eksitus, 2026-07-24
Source: @farawaythrower, 2026-07-25
Source: @jonathanyjoerrington, 2026-07-27
Source: @btakk_ak47, 2026-07-27
Source: @aifonder_66800, 2026-07-29
Source: repo github.com/studebaker8/cmp170hx-gen2 by studebaker8 on GitHub (work by bendy2, studebaker8, amoghmunikote and abobasixseven; lab sweep 2026-08-06)
Source: repo github.com/abobasixseven/unlock-cmp-170hx by abobasixseven on GitHub (lab sweep 2026-08-06)
Source: @tchelobh, 2026-08-20
Source: @olehborzykh, 2026-09-01
Source: @dottoreargus, 2026-09-07
Source: repo github.com/rjohny55/cmp170hx-unlocker by rjohny55 on GitHub (lab sweep 2026-09-24)
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub (lab sweep 2026-09-24)
Source: @darkmaniac7, 2026-09-24 (lab notes)
Source: repo github.com/PixelML/club-170hx by PixelML on GitHub (docs/CMP170HX-UNDERSTANDING.md, posted in Discord by @c133)
Gen2 depends on the board: the same cards work on consumer boards and fail on an ASUS X99-A and a Z10PA-D8
Jonathany Joerrington, on an ASUS X99-A (LGA2011) with virtualization enabled as the guide says, got at best PCIe 2.0 x1, and only in one specific slot of four, or else PCIe 1.0 x4. btakk_ak47 got Gen2 on a consumer board, but it failed on an ASUS Z10PA-D8 (he wrote "Z10PAD8"; X99 era). aifonder saw the same split between a consumer board and his W890 SAGE (below). ju5t5omedude suspected a too-new kernel; Studebaker said it is not the kernel.
Reported
Source: @jonathanyjoerrington, 2026-07-27
Source: @btakk_ak47, 2026-07-27
Source: @ju5t5omedude, 2026-07-27
Source: @_studebaker_, 2026-07-27
Results by board
Bare-metal hosts, one row per report. Empty cells were not given in the post. The result is the link as reported; width depends on the slot and the capacitor mod, not on the unlock. The sections below hold the detail, the status and every source for each row.
| Board | CPU | OS | Kernel | Driver | Result | Credit |
|---|---|---|---|---|---|---|
| AMD host (board not given) | AMD | HiveOS (Ubuntu 22.04) | 6.12.0 | Gen2 x4 on two cards with Newman's multi-card retrain fork; no capacitor mod, no grub change | Source: @newman_08462, 2026-07-24 | |
| HiveOS host (board not given) | HiveOS | The plain install commands did not work | Source: @daniyel0042, 2026-07-30 | |||
| ASUS Prime Z370 | Intel i3-8100 | Gen2 on two 10 GB cards with Newman's fork (the early branch left both at Gen1); the x16 capacitor-modded card trains x8, not x16 | Source: @gmzombie, 2026-07-24 | |||
| Supermicro H12DSi-N6 | 2x AMD EPYC 7713 | NixOS | 6.18.38 | Gen2 on all three 10 GB cards (40 GB), patches applied in the kernel build | Source: @eksitus, 2026-07-24 | |
| Consumer board with one x16 slot (model not given) | Gen2 with bendy2's repo | Source: @aifonder_66800, 2026-07-29 | ||||
| ASUS "W890 SAGE" (the owner's name for it) | Intel ("676x") | Ubuntu 24.04 | 7.0.0-28-generic | Fails: LnkCap Gen2 but LnkSta Gen1 (capacitor-modded cards Gen1 x16, unmodded Gen1 x4); black screen after the script | Source: @aifonder_66800, 2026-07-24 | |
| ASUS ROG Rampage VI Extreme Omega | Intel (X299 class) | Arch | 7.1.3 | Fails: the card advertises Gen2 but does not train to it; unresolved | Source: @farawaythrower, 2026-07-25 | |
| Board not given (Plean) | Ubuntu 22.04 | 5.15.0-186, later 6.12 | open 610.43.02 | Gen2 x4 on an 8 GB card at 64 GB; failed first on Ubuntu 26.06 with kernel 7.0.0 | Source: @plean., 2026-07-26 | |
| ASUS X99-A (LGA2011) | Intel | Fails: at best Gen2 x1 in one slot of four, otherwise Gen1 x4 | Source: @jonathanyjoerrington, 2026-07-27 | |||
| ASUS Z10PA-D8 | Intel Xeon E5 class | Fails; the same owner reaches Gen2 on a consumer board | Source: @btakk_ak47, 2026-07-27 | |||
| ASUS Pro WS W790E-SAGE SE | Intel Xeon w5-3433 (one post says w5-3423) | Width, not speed: 3 of 5 x16-capable cards trained x8 in x16 slots; cause unknown | Source: @diabeetee, 2026-07-22 | |||
| AMD A10-7800 host | AMD A10-7800 | The unlock fails; the same card works on another host with Ubuntu 26 | Source: @xeax, 2026-08-01 | |||
| X99 / C612 board (model not given) | Intel Xeon E5 | Fails: the board trains the card at Gen1 at power-on and no retrain catches it | Source: @tchelobh, 2026-08-20 | |||
| Debian 13 / Proxmox host, card on the host (with a CMP 90HX) | Debian 13 / Proxmox | 6.8.12-15-pve | combined 170HX + 90HX build of open 610.43.03 | Gen2 x4 on a 10 GB card (40 GB) once pcie-gen2.patch was in the build | Source: @olehborzykh, 2026-09-01 | |
| Z390 board with one PEX8747 switch | Intel (Z390 class) | Gen2 behind the switch with a boot-time resync in a strict order; one card | Source: @peter643, 2026-09-04 | |||
| ASUS Prime TRX40-Pro (BIOS 2502) | AMD Threadripper 3970X | 610.57.04 (amogh upstream) | All slots unlock once a leftover pci=nocrs was removed from GRUB; the top slot ran 64 GB at Gen2 x16 | Source: @nestaa51, 2026-09-19 | ||
| EPYC Rome system (board not given) | AMD EPYC Rome | The unlock works on all 8 cards (link speed not stated) | Source: @cute_sc2, 2026-09-19 | |||
| TRX50 system (board not given) | AMD Threadripper (TRX50) | 2 cards with no problems (link speed not stated) | Source: @cute_sc2, 2026-09-19 | |||
| Single-socket Xeon E5 v4 host | Intel Xeon E5 v4 | Gen1 x16 to Gen2 x16 after a function-level reset and a second retrain pass | Source: repo github.com/rjohny55/cmp170hx-unlocker by rjohny55 on GitHub | |||
| AMD EPYC host (board not given) | AMD EPYC | 615, with a probe-serialization patch | Eight cards at Gen2 x16 on the first pass, all keeping 65536 MiB | Source: repo github.com/rjohny55/cmp170hx-unlocker by rjohny55 on GitHub | ||
| Reproduction host (board not given) | Debian 13 | 6.12.96 | open 610.43.03 with the early patches 0001 to 0008 | Gen1 x4 to Gen2 x4 on one 8 GB card | Source: repo github.com/luannanxian/cmp170hx-gen2 by luannanxian on GitHub |
Hosts where Gen2 trains
AMD with HiveOS (Ubuntu 22.04), kernel 6.12.0: Gen2 x4 on two cards, no mod and no grub change
Newman's working Gen2 setup: an AMD CPU, HiveOS based on Ubuntu 22.04, kernel 6.12.0, no capacitor mod and no change to the grub command line. Two 170HX ran at Gen2 x4 after his multi-card retrain changes, the fork GMZombie007 and others used when the early branch failed with two cards. A second HiveOS host failed with the plain install commands (see HiveOS below).
Reported
Source: @newman_08462, 2026-07-24
ASUS Prime Z370 with an i3-8100: Gen2 on two 10 GB cards; the capacitor-modded card reaches only x8
GMZombie007's test rig: ASUS Prime Z370, i3-8100, 8 GB RAM. With Amogh's early Gen2 branch his two 10 GB cards stayed at 2.5 GT/s and his x16-modded card showed x4. With a single card, sudo ./retrain.sh gave 5.0 GT/s, but with both cards the retrain errored on the first GPU; Newman's fork worked with both. At Gen2 the x16 capacitor-modded card came up at x8, not x16, and an OpenCL benchmark showed about 3.24 GB/s over PCIe against 1.67 GB/s on the unmodded card.
Reported 8 GB and 10 GB
Source: @gmzombie, 2026-07-24
Supermicro H12DSi-N6, 2x EPYC 7713, NixOS kernel 6.18.38: Gen2 on three 10 GB cards
Eksitus runs a Supermicro H12DSi-N6 with two Zen 3 EPYC 7713 on NixOS kernel 6.18.38 and applies the cmpunlocker patches in his kernel build instead of using the .sh scripts. With Newman's 0008 patch pulled in, Gen2 works on all three of his 10 GB cards unlocked to 40 GB. The motherboard slots as they are cap him at 3 cards; he planned SlimSAS adapters for 8 or more cards at x8. On one of these cards he measured what Gen2 buys with the model fully in VRAM: Qwen3.6-27B UD-Q8_K_XL on ik_llama with MTP, Gen2 x4 against Gen1 x4, everything else unchanged. Prefill: pp512 277.8 against 203.8 tok/s, pp2048 449.4 against 328.8, pp8192 493.9 against 363.3 (about 36% more). Decode: tg128 41.5 against 38.2 tok/s, tg512 40.1 against 37.7, tg2048 37.9 against 36.8 (3 to 9% more). At Gen1, enabling MTP cost a lot of prefill; at Gen2 with MTP, prefill came close to Gen1 without MTP. Why link speed matters with the whole model on the card is unexplained: Eksitus guessed MTP needs CPU coordination and lower latency, while Arcanum thought MTP should run entirely on the GPU.
Reported 10 GB
Source: @eksitus, 2026-07-24
Source: @eksitus, 2026-07-27
Source: @_arcanum_, 2026-07-27
Plean: Gen2 x4 on an 8 GB card with Ubuntu 22.04, kernel 5.15 and a plain install
Plean got Gen2 x4 on an 8 GB card unlocked to 64 GB using the Gen2 branch; he never touched main. His steps: enable VT-d in the BIOS, install Linux fresh, install the NVIDIA open 610 driver, reboot, check nvidia-smi, run ./install.sh. No uninstall and no manual IOMMU flags: the script handled IOMMU. It first failed on Ubuntu 26.06 with kernel 7.0.0, then worked on Ubuntu 22.04 with kernel 5.15.0-186 and driver 610.43.02, and later also on kernel 6.12.
Reported 8 GB
Source: @plean., 2026-07-26
ASUS Prime TRX40-Pro with a 3970X: slots 2 and 3 failed GSP Booter Load until a leftover pci=nocrs was removed
On an ASUS Prime TRX40-Pro (BIOS 2502) with a Threadripper 3970X, Nestaa51 ran amogh upstream at driver 610.57.04 (cmpunlocker2 also failed): only the top slot unlocked (64 GB, Gen2 x16). Slots 2 and 3 failed with "kgspBootstrap_TU102: failed to execute Booter Load (ucode for initial boot): 0xffff", "Max GSP-RM boot attempts exceeded: 4/4" and "RmInitAdapter failed! (0x62:0xffff:2119)", and the failed init often hard-locked the box a minute later (2026-09-19). AER was clean and BAR1 still mapped 64 GB. Above 4G decoding and ReBAR on, 10-bit tags off, the preferred IO slot set to slot 2, pci=realloc with hpmmioprefsize up to 8T, the gen2 service on or off, a build without the pcie-gen2 patch and cold boots did not help. The cause was pci=nocrs left on the GRUB command line from his earlier Intel setup, which broke PCI slot ID assignment on the Threadripper board; removing it made the unlock work in all slots. giaki3003's rule of thumb: regenerate GRUB and the kernels after switching platforms or microcode. CUTE was in the thread.
Reported 8 GB
Source: @nestaa51, 2026-09-19
Source: @cute_sc2, 2026-09-19
Source: @giaki3003, 2026-09-19
EPYC Rome unlocks all 8 cards; TRX50 runs 2 cards without problems
CUTE reported an EPYC Rome system where all 8 cards unlock, plus 2 cards in a TRX50 system with no problems (2026-09-19). Nestaa51, comparing with his TRX40 problem, noted that Rome shares the Zen 2 IO die with the 3970X. Morrowmake asked whether some of the Rome cards run at x8 or behind PCIe switches; there was no answer, and the link speed on these hosts was not stated.
Reported
Source: @nestaa51, 2026-09-19
Source: @cute_sc2, 2026-09-19
Source: @voidgiraffe, 2026-09-20
Debian 13 / Proxmox host with a 170HX and a CMP 90HX: Gen2 only once pcie-gen2.patch was in the combined driver
oleh borzykh runs a 10 GB card ("CMP 170HX 40GB", 10de:2082) and a CMP 90HX in one Debian 13 / Proxmox host, kernel 6.8.12-15-pve, with his own combined build of the NVIDIA open driver 610.43.03 for both cards. The cards run on the host itself, not in a guest. The Gen2 watcher from cmpunlocker PR #35 ran in both the initramfs and the systemd stage but never saw a change: LnkSta stayed 0x1041 (Gen1 x4). kmike's diagnosis: pcie-gen2.patch had never been included in the combined build. With it merged and a reboot, the watcher caught the change 6.262 s after load and retrained to 0x1042 (Gen2 x4), and both the 40 GB unlock and the 90HX compute unlock kept working. He said the unlock is plain cmpunlocker and his problem was the patch steps. His combined build steps wait for the author; the 90HX side is on CMP 90HX unlock hub.
Reported 10 GB
Source: @olehborzykh, 2026-09-01
Source: @kmike84, 2026-09-01
Source: @mikevlz_59313, 2026-09-01
Z390 board with one PEX8747 switch: Gen2 needs a boot-time resync in a strict order
peter643 got a CMP 170HX running at Gen2 behind the single PEX8747 switch on his Z390 board. The problem was timing: the Gen2 resync has to run at boot, and the operations must happen in a very specific order. He offered to post his steps; they are not in the history, and the sequence waits for the author. One PEX8747 gives at most 4 links at x8 Gen2. He did not test boards with two PEX8747 switches, where the cards sit in two separate pairs and not every card can reach every other card without going through the CPU. As of 2026-09-05 he had only one card, so P2P through the switch is unverified. More on these boards: PCIe switch boards (PEX88096, PEX88064, PEX8747, PLX8749).
Reported
Source: @peter643, 2026-09-04
Source: @cressida1881, 2026-09-04
luannanxian/cmp170hx-gen2: Gen1 x4 to Gen2 x4 on one 8 GB card, Debian 13, kernel 6.12.96
A Chinese-language reproduction repository for one 8 GB card (10de:20c2) at x4, based on amoghmunikote/cmpunlocker at a4de322 (summarized from Chinese). The early patches 0001 to 0008 on open driver 610.43.03 make the card advertise Gen2, but the link stays at Gen1 until the card's direct root port retrains. Verified on Debian 13 with kernel 6.12.96. The repo reports llama.cpp Qwen3.6-27B going from Gen1 x4 to Gen2 x4 at +33.5% to +35.7% prefill on short prompts and +5.6% to +7.2% decode. Historical: it carries only the eight early patches, driver 610.43.03 and a single card.
Reported 8 GB
Source: repo github.com/luannanxian/cmp170hx-gen2 by luannanxian on GitHub (posted in Discord by @aaasss1456)
Hosts where Gen2 fails or the link is short
ASUS W890 SAGE (Intel): Gen2 fails, black screen after the script; a consumer board works
aifonder ran two 170HX directly in an ASUS board he calls W890 SAGE (4 full PCIe 5 x16 slots, a "676x" CPU) on Ubuntu 24.04, kernel 7.0.0-28-generic. After the Gen2 script LnkCap showed 5GT/s but LnkSta stayed at 2.5GT/s: his x16 capacitor-modded cards trained Gen1 x16 and an unmodded card Gen1 x4. The grub line "quiet splash intel_iommu=on iommu=pt iomem=relaxed" did not help, Newman's fork also left the unmodded card at 2.5GT/s x4, and cards in the first and fourth slots both stayed at 2.5GT/s. On 2026-07-29 the unlock script left the system unable to load the GUI (black screen) with LnkCap Gen2 and LnkSta Gen1; the same day he reported Gen2 working on a consumer board, which has only one x16 slot, using "bendy2's repo". btakk_ak47 suggested the unlock process may still have a bug. No fix was posted. Luna said the board did not match its box and named it a W890-SAGE; aifonder uses the non-SE version, and the CPU model is not confirmed.
Reported
Source: @aifonder_66800, 2026-07-22
Source: @lunaroseoc, 2026-07-22
Source: @aifonder_66800, 2026-07-24
Source: @aifonder_66800, 2026-07-27
Source: @aifonder_66800, 2026-07-29
Source: @btakk_ak47, 2026-07-29
ASUS ROG Rampage VI Extreme Omega (Intel), Arch, kernel 7.1.3: Gen2 not negotiated
Far could not get Gen2 to negotiate on an 8 GB card: the card advertised Gen2 but did not train to it. His Rampage VI Extreme Omega has no explicit IOMMU option in the BIOS, but VT-d was enabled; he set the IOMMU kernel parameters by hand on Arch (kernel 7.1.3) and reinstalled cleanly (uninstall, then install). He posted his dmesg output and was pointed to the setup channel. Unresolved.
Reported 8 GB
Source: @farawaythrower, 2026-07-25
Source: @robbbbbb7973, 2026-07-25
Source: @_j3ns, 2026-07-25
X99 / C612 boards: the card trains at Gen1 before the OS loads, and no retrain catches it
tchelobh could not get the Gen2 unlock working on an old X99 board and blamed quirks of the Xeon E5 PCIe controller (2026-08-20). After three days on an X99 C612 board he found it trains the card at PCIe 1.0 right at power-on, and no retrain script or kernel patch retrains it. Forcing Gen2 in the BIOS does not help, because before the OS loads the card advertises only Gen1, which changes only after boot with the unlocker (2026-08-26). ju5t5omedude understood the retrain to happen when the driver loads with the unlock, and giaki3003 discussed the timing. tchelobh considered patching the board BIOS or a card VBIOS that exposes Gen2 natively; CUTE said someone was working on a custom VBIOS but it might never happen. His use is OCR, where Gen2 matters more than for decode. Not every X99 board fails: SeriousRising runs Gen2 cards on an X99 board under Proxmox (see Proxmox Gen2 passthrough). See also Known bad combinations.
Reported
Source: @tchelobh, 2026-08-20
Source: @giaki3003, 2026-08-26
Source: @cute_sc2, 2026-08-26
Source: @ju5t5omedude, 2026-08-26
AMD A10-7800: the unlock fails; the same card works on another host
Xeax's unlock failed on an AMD A10-7800 platform, while the same card worked on another platform running Ubuntu 26. He attached the failure log and a boot-time error log. No resolution was posted.
Reported
Source: @xeax, 2026-08-01
ASUS Pro WS W790E-SAGE SE: x16-modded cards train at x8 in x16 slots
diabeetee's board is an ASUS Pro WS W790E-SAGE SE with an Intel Xeon w5-3433 (112 lanes); a 2026-07-22 post names a w5-3423, so the two posts disagree on the CPU. With 5 cards fitted, the cards reported x16 capable but 3 ran at x8, although the BIOS showed all slots but one as x16; the cause is unknown. Two cards with the x16 capacitor mod both show the right LnkCap, but one reads x8 in LnkSta, and both sit in x16 slots (W-3400 CPUs give x16/x16/x16/x16/x16/x8/x16). Suggested checks: swap the cards to see whether the fault follows the card or the slot, and confirm in the BIOS (Advanced, Onboard Devices Configuration) that the slots are set to x16 (zapirate). Danske suggested BIOS settings or a bad adapter, and ju5t5omedude suggested moving cards around and said it might be the patch. Unresolved. By 2026-08-01 he had fitted 7 cards in a server cabinet and planned 12 to 13; models load from a Gen4 NVMe at 3 to 4 GB/s, in about a minute.
Reported
Source: @diabeetee, 2026-07-22
Source: @diabeetee, 2026-07-30
Source: @meromgreen, 2026-07-30
Source: @zapirate, 2026-07-30
Source: @danske1, 2026-07-30
Source: @meatsus, 2026-07-31
Source: @ju5t5omedude, 2026-07-31
Kernels and operating systems
More on choosing an OS and kernel: Operating systems and kernels.
Kernel version and Gen2: works on 5.15, 6.12 and 6.18.38; 7.0.0 failed for some
Reports disagree on whether the kernel matters; the sides in the order they were posted:
- Newman's working setup was AMD with HiveOS (Ubuntu 22.04) on kernel 6.12.0. Source: @newman_08462, 2026-07-24
- Studebaker rolled CachyOS back to 6.12-LTS and concluded the kernel was not the cause. Source: @_studebaker_, 2026-07-24
- Plean failed on Ubuntu 26.06 with kernel 7.0.0, then succeeded on a fresh Ubuntu 22.04 (kernel 5.15) and on 6.12; swapping mainline kernels on Ubuntu caused trouble with the 610 open driver. Source: @plean., 2026-07-26
- Eksitus confirmed 6.18.38 on NixOS. Source: @eksitus, 2026-07-26
- ju5t5omedude stayed at Gen1 on the latest Ubuntu 7.x kernel after trying Amogh's and Newman's patches. Source: @ju5t5omedude, 2026-07-26
- ProfDD asked whether only 5.15 works and said 7.0.x does not; Plean rejected that ("I've gotten 5.15 and 6.12 working for me"), and giaki3003 told him to use the new script. Source: @profdd, 2026-07-26
Source: @giaki3003, 2026-07-26 - Far tested on 7.1.3 and thinks it is not kernel dependent (his own card did not negotiate Gen2 on that kernel; see the Rampage VI entry above). Source: @farawaythrower, 2026-07-26
Contested 8 GB and 10 GB
Ubuntu 26.04: Gen2 works on some boards, not others
The sides in the order they were posted:
- Plean's first attempt, on Ubuntu 26.06 with kernel 7.0.0, failed; Ubuntu 22.04 then worked (see his entry above). Source: @plean., 2026-07-26
- CUTE: the Gen2 unlock did not work on Ubuntu 26.04, only on 24.04. Source: @cute_sc2, 2026-08-26
- Ryko: it worked on 26.04 for him. ju5t5omedude thinks he also runs 26.04; his Gen2 unlock took several tries in the OS before it stuck, and he could see the link speed change in GPU-Z. Source: @vektor_prime, 2026-08-26
Source: @ju5t5omedude, 2026-08-26 - Merom Green and bayley_w confirmed Gen2 on Ubuntu 26 (bayley_w: allreduce at 1.6 GB/s). Jonathany measured Gen2 x4 at 1.63 GB/s host to device and 1.66 GB/s device to host, 94% of the about 1.76 GB/s expected. Source: @meromgreen, 2026-09-07
Source: @jonathanyjoerrington, 2026-09-07
Source: @bayley_w, 2026-09-07 - tchelobh's setup did not work; CUTE saw it fail on one mainboard but work on another with Ubuntu 26; vskayp had x16 Gen2 on Ubuntu 24 with the same card. Source: @tchelobh, 2026-09-07
Source: @cute_sc2, 2026-09-07
Source: @vskayp, 2026-09-07
dottoreargus noted in the same thread that nvidia-smi shows the device maximum as Gen1 while the unlocker negotiates Gen2, so judge the result by LnkSta. Source: @dottoreargus, 2026-09-07
Contested
HiveOS: one working Gen2 setup, one failed install
- Newman ran Gen2 on HiveOS based on Ubuntu 22.04, kernel 6.12.0: two cards at Gen2 x4 on an AMD CPU, no grub change, with his own fork (2026-07-24). Source: @newman_08462, 2026-07-24
- Asked by Daniyel whether the unlocker works on HiveOS, btakk_ak47 answered "Can", but Daniyel later said running the install commands on HiveOS did not work, with no follow-up (2026-07-30). Source: @daniyel0042, 2026-07-30
Source: @btakk_ak47, 2026-07-30
The difference (Newman's own fork against the plain install commands) was not discussed. mfj4 asked whether anyone runs the unlock under WSL2; nobody answered. Source: @mfj4, 2026-07-31
Contested
In a virtual machine
A guest usually gets the memory unlock but not Gen2, because the guest cannot retrain the physical port. Methods that do reach Gen2 in a Proxmox guest by retraining from the host are on Proxmox Gen2 passthrough; general passthrough setup is on Virtual machines and passthrough.
| Host | CPU | OS | Kernel | Driver | Result | Credit |
|---|---|---|---|---|---|---|
| Proxmox hosts (July and August 2026) | Proxmox VE with a Linux guest | The memory and compute unlock work in the guest; Gen2 does not | Source: @jimtjames, 2026-08-03 | |||
| Dell server with iDRAC (model not given) | Proxmox | older unlocker code with the Gen2 patch removed; later the SM patch | Gen1 only; Gen2, and possibly ReBAR, not working in the VM | Source: @ikyo, 2026-09-09 | ||
| Proxmox host with two 8 GB cards (board not given) | Proxmox VFIO guest with SeaBIOS | Gen1 x4 with the Gen2 patches dropped; no P2P | Source: @marklintonca, 2026-09-17 | |||
| Dell PowerEdge R730 | 2x Intel Xeon E5-2695 v4 | Proxmox VE 9.1.1 host, Ubuntu 24.04.4 guest | 6.17.2-1-pve (host), 6.8.0-142-generic (guest) | modified open 610.43.03 | Gen2 x4 in the guest on one 8 GB card at x4, after one timed retrain from the host | Source: @mazzanfar7661, 2026-09-29 |
A VM gets the memory unlock but not the Gen2 link: the guest cannot retrain the physical port (July to August 2026)
On 2026-07-30 there was no Gen2-in-VM support; Amogh said the only issue for any passthrough is the Gen2 retraining. Humvee55 said it works better with a host-side retrain and depends on the hypervisor, and Capone said the Proxmox host must retrain the link, unlike bare metal. On 2026-08-02 gym (jimtjames) reported that the compute and memory unlock works inside a VM but the Gen2 unlock does not, while the full unlock works on bare metal; robbbbbb7973 agreed that Proxmox does not work for Gen2 and relayed another member's claim that it worked under ESXi passthrough. Capone has used the unlock inside a Proxmox VE VM. On 2026-08-19 ppc64_ hit the same Gen2 problem under Proxmox PCIe passthrough. replacedriver47 proposed LXC instead: the Proxmox host runs the patched NVIDIA driver with its kernel modules, and the container gets the matching userspace-only driver; Amogh said LXC should work but needs work. Aeiou found TrueNAS Scale handled the GPU better than some dedicated hypervisors. Ikyo asked whether one card could be split across several VMs; nobody answered.
Reported 8 GB
Source: @._.patrik, 2026-07-30
Source: @_j3ns, 2026-07-30
Source: @scallion_9883, 2026-07-30
Source: @beefbonecapone, 2026-07-30
Source: @replacedriver47, 2026-07-30
Source: @_____josh_____, 2026-07-30
Source: @ikyo, 2026-08-02
Source: @91d5a66bffc978331fac75b68fa86aba, 2026-08-02
Source: @aaasss1456, 2026-08-02
Source: @beefbonecapone, 2026-08-02
Source: @robbbbbb7973, 2026-08-02
Source: @jimtjames, 2026-08-03
Source: @_j3ns, 2026-08-03
Source: @ppc64_, 2026-08-19
Dell server with Proxmox: Gen1 only after removing the Gen2 patch; Gen2 and ReBAR unsolved
Ikyo could only get his cards working at the base Gen1 link on a Dell server (iDRAC) running Proxmox: he had to strip the unlocker's Gen2 patch completely (on an older version of the code), and other forks did not work either; he suspected the Dell platform and ReBAR (2026-09-09). Other devices on the host, SFP28 network cards, linked at 8GT/s x8. Amogh asked whether other devices train above Gen1 and moved the case to a support ticket. By 2026-09-22 he had tested Proxmox with Amogh and applied the SM patch, but Gen2, and possibly ReBAR, still did not work in the VM (host with 256 GB RAM). Waka324 advised plain Ubuntu Server for a single node (Proxmox gives more ways to break things), and segmond suggested Ubuntu with Linux containers.
Reported
Source: @ikyo, 2026-09-09
Source: @_j3ns, 2026-09-10
Source: @cmdrwaka324, 2026-09-22
Source: @segmond, 2026-09-22
Proxmox VFIO with two 8 GB cards: SeaBIOS guest, Gen2 patches dropped, Gen1 x4
Mark Linton runs two 8 GB cards, 64 GB each after the unlock, in a Proxmox VM through VFIO. The guest has to boot with SeaBIOS: UEFI guest firmware breaks the unlock (2026-09-17). He dropped the Gen2 patches for passthrough, so the link is PCIe Gen1 x4, with no NVLink or P2P; Zorg33 asked why Gen1.
Pointer: a longer write-up by @marklintonca (the VM configuration and a VBIOS reflash of both cards) exists; see Discord. The details wait for the author.
Reported 8 GB
Source: @marklintonca, 2026-09-17
Source: @zorg33, 2026-09-17
Dell PowerEdge R730 under Proxmox: Gen2 x4 in the guest after one timed host-side retrain (one-card experiment)
mazzanfar posted his repository on 2026-09-29 as a guide and configuration that worked for Proxmox users with Gen2 problems; cibernox replied he had the same issue until he started using a PCIe switch. Tested setup: a Dell PowerEdge R730 with two Xeon E5-2695 v4, Proxmox VE 9.1.1 on kernel 6.17.2-1-pve, an Ubuntu 24.04.4 guest on 6.8.0-142-generic with Q35 and SeaBIOS, one 8 GB card (10de:20c2) unlocked to 65,536 MiB on a physical x4 link, a modified nvidia-open 610.43.03 and a 150 W power limit. A watcher on the host polls the card and retrains the root port once, at the moment the guest driver asks for Gen2 during its init; the link then read Gen2 x4 at both ends, and host transfers roughly doubled. No inference speedup was measured. Gen2 x4 survived one VM stop and start, but a host reboot and an AC power cycle were not tested, the packaged watcher never ran on hardware, it accepts only 10de:20c2, it needs the root port already targeting Gen2, and the build targets driver 610.43.03 rather than current mainline. Licence GPL-2.0-only.
Pointer: the repository's register details, commands and test numbers wait for the author; see Discord and the repository.
Reported 8 GB
Source: @mazzanfar7661, 2026-09-29
Source: @cibernox, 2026-09-29
Source: repo github.com/mazzanfar/cmp170hx-proxmox-pcie-gen2 by mazzanfar on GitHub (last commit cb8d757, 2026-09-29, posted by @mazzanfar7661)
Related
Gen2 link retrain, Verifying an unlock, Proxmox Gen2 passthrough, Virtual machines and passthrough, Operating systems and kernels, Kernel parameters and module options, Known bad combinations, PCIe x16 capacitor mod, PCIe switch boards (PEX88096, PEX88064, PEX8747, PLX8749), CMP 90HX unlock hub, Dell PowerEdge R730, All platforms, Platforms.