Cmpunlocker forks

cmpunlocker (amoghmunikote) is the mainline. Several forks exist, each adding one thing the mainline lacks. All are GPL-2.0 and build the same way (download the stock NVIDIA open modules, apply patches, build locally). Pick by what you need; do not stack them.

Morrowmake/cmpunlocker (PCIe P2P)

A fork of mainline (base 6c442ee) that adds working GPU-to-GPU P2P on the CMP 170HX. It carries the maintainer's p2p-unlock.patch from the upstream P2P branch and opens TRAP31_PLM through the Booter only when ForceP2P is set. It is the ready-to-install P2P driver that Morrowmake's GLM-5.3-Flash and DeepSeek-V4.1-Flash recipes require. Source: @mountains1605, 2026-10-20

Install: sudo ./install.sh --p2p --profile=8gb --no-iommu --no-passthrough (use

 --profile=10gb for 10 GB cards; drop the --no- flags to keep those setups). It writes
 options nvidia NVreg_RegistryDwords="RmForceEnableGen2=1;RMPcieLinkSpeed=0x1;ForceP2P=0x11" and keeps
 ForceP2P=0x11 on later installs.

Supported drivers (driver/VERSION): 615.71.09, 610.57.04, 610.43.03, 610.43.02. It does not carry

 the ECC patches (removes ecc-enable, ecc-fbpa-static, ecc-reporting) and does
 not list 615.71.10 or 615.78.08, so it is behind mainline.

Verify with python3 tools/p2p-content-check.py; nvidia-smi topo alone does not prove the

 bytes arrive. Measured about 6.65 GB/s each way on 16 MiB copies, 4 cards, 610.57.04, Gen2 x16.

P2P only works where trap 31's PLM reads 0xFFFFFFFF (writable); other parts stay unsupported by design.
Origin: curiouspixel's P2P patch in this fork fixed P2P for Trust Inter and GhostInTheDatacenter on 2026-10-07.

 Source: @trustinter, 2026-10-07

Repo: Morrowmake/cmpunlocker. See P2P for the full mechanism.

vektorprime/cmpunlocker (170tune memory lever)

Master is identical to mainline; the fork exists for its branch fbpa-pll-170tune (merged 2026-10-07). It keeps the mainline unlock and adds the FBPA_MEM, broadcast FBPA_PLL0/1/2 and unicast FBPA0 PLL (0x00903c7c) privilege-mask opens from the asm64 lineage, deliberately without baking a memory clock, so the card boots stock and the paired vektorprime/170tune (branch fbpa-first-active) can read and write the HBM PLL over BAR0 at runtime. On the 10 GB card the unicast FBPA0 row is skipped (that window is orphaned in hardware and faults GSP if forced); the broadcast PLLs still open. Source: repo github.com/vektorprime/cmpunlocker by vektorprime on GitHub, posted by @rh3dai, 2026-10-04
Source: @vektor_prime, 2026-10-07 Repo: vektorprime/cmpunlocker (branch fbpa-pll-170tune), by @vektor_prime (Ryko). Credit master to amoghmunikote.

asm64-hooligan/cmpunlocker (the original "unlobotomize" tool)

The asm64 lineage: patched 610.43.03 / 610.43.02 modules, all unlock logic in one file (driver/src/cmpunlock.c). Its distinguishing additions are a compile-time HBM PLL overclock (--mclk-ndiv=N, clock = N x 27 MHz), DRAM timing scaling (--mclk-timings), live BAR0 timing tools and anti-rollback persistence across kernel updates. Mailbox --p2p runs in a 512 KB window inside the stock 64 MB BAR1 (no ReBAR, no Above 4G needed), about 1.7 GB/s between GPUs behind one switch at Gen2 x4. Stock NDIV table: 10 GB any VBIOS 45 (1215 MHz), 8 GB 250 W VBIOS 54 (1458 MHz), 8 GB 300 W VBIOS 64 (1728 MHz). It is pinned to 610.43.0x and has no ECC, so it is behind mainline for most users; it is the base cachenetics/170tune's memory tooling credits. Source: @c133, 2026-10-22 Repo: asm64-hooligan/cmpunlocker (commit author "aboba").

Warning The repo ships a prebuilt x86-64 benchmark binary; do not run untrusted binaries. Memory overclock can corrupt silently; the repo's own rule is zero gpu-burn errors over 300 s before trusting a value. --mclk-timings tCCD_L = 3 hard-hangs the controller and needs a reboot to recover.

AnalyticETH/cmpunlocker (BAR1 work, repo gone)

A personal fork where the author said he had made "messy, beta" progress on the BAR1 / ReBAR side on an ASUS W680-ACE board. Source: @analyticeth, 2026-10-07 As of 2026-10-30 the repository returns 404 on github.com, so do not link it. The BAR1 patches that reached the community live in mainline cmpunlocker. Recorded here only as the historical origin of the BAR1 attempts.

Related consumer-card tools (not for the 170HX)

These descend from the same SEC2/PLM/SS0-SS1 technique but their device tables do not include GA100, so they do not unlock the CMP 170HX. They belong here as sibling projects.

Zagitta/gpu-unlock (link): a Linux kernel module that removes NVIDIA's

 half-rate tensor cap on GeForce RTX 20/30/40 cards, at boot, with no driver patching. Its GA100 PLM table is copied
 from cmpunlocker commit 2d10f891 but no GA100 card IDs are registered, so CMP cards are refused. Zagitta is the author
 of the consumer port that Tom's Hardware covered on 2026-10-26.
 Source: @zagitta, 2026-10-18
 Zagitta on 2026-10-25: the unlock targets 30-series cards and leaves other GPUs in the host alone; it covers GA102,
 GA104 and GA106. Measured gains and its lineage are on Tools#Related work on other NVIDIA cards.
 Source: @zagitta, 2026-10-24

 Source: @zagitta, 2026-10-25

JustAResearcher/gpu-unlock-ada (link): the same tensor

 unlock for RTX 40-series (AD102, AD103), derived from Zagitta/gpu-unlock; GPL-2.0-only. Posted with a report of an
 unlocked RTX 4090. Not for the 170HX. Details on Tools#Related work on other NVIDIA cards.
 Source: repo github.com/JustAResearcher/gpu-unlock-ada by JustAResearcher on GitHub, posted by @petrikrohn_26497, 2026-10-24

dartraiden/NVIDIA-patcher (link): patches the stock driver so

 mining cards (including the 170HX, both IDs) bind as display/3D GPUs. On Linux linux.sh hex-patches
 nv-kernel.o_binary in the official .run. It is not an unlock (no FP32 FMA, memory or
 PCIe change) and patches the proprietary module, so it is an alternative to cmpunlocker for Windows/graphics use, not
 a companion. Source: @xeax, 2026-10-27

mintoriakamoto/CMP100-210xFULLTENSORSxGEN3x1: a Tensor + Gen3 unlock for the CMP 100-210 (GV100 Volta),

 not the 170HX. The author was flagged as untrustworthy (banned on the server after a PCIe 5.0 claim); treat with
 caution. Source: @_j3ns, 2026-10-09

Fork comparison

Fork	Adds	Has ECC?	Drivers	Best for
mainline (amoghmunikote)	the baseline	yes	610.43.02 to 615.78.08	almost everyone; 8 GB cards
Morrowmake	mailbox + ForceP2P P2P	no	610.43.02 to 615.71.09	multi-card P2P recipes
vektorprime (branch)	170tune HBM PLL lever	yes (from master)	610.43.02 to 615.71.09	memory tuning with 170tune
asm64-hooligan	`--mclk-ndiv` OC, anti-rollback	no	610.43.0x only	HBM overclock, kernel-update persistence

For P2P-and-ECC in one package, there is no release yet; a feature request to fold P2P into mainline as v0.6 is open. Source: @rh3dai, 2026-10-23