cmpunlocker v0.5, the mainline driver unlock for the CMP 170HX, was released on October 4, 2026, per its maintainer @_j3ns. It takes the 8 GB card to 64 GB and the 10 GB card to 40 GB, adds 4 SMs on both and turns on ECC; the Driver unlock page lists it as the current release.
What v0.5 gives you
It patches NVIDIA's open kernel modules. Nothing is flashed or burned, and a boot service reapplies the unlock after every power cycle.
- 8 GB card: 65536 MiB (64 GB), 74 SMs instead of 70, PCIe Gen2, ECC reporting and a 64 GB BAR1.
- 10 GB card: 40960 MiB (40 GB) and the same 4 extra SMs.
@68x0 reproduced the release on master the same evening it shipped, independently of the maintainer (confirmation). On that basis the Driver unlock page marks the core unlock Proven, meaning a second person reproduced it.
What it does not do
As of master commit 2c7e2ec (October 27, 2026), cmpunlocker does not unlock PCIe Gen3 or Gen4, the 10 GB card beyond 40 GB, more than 74 SMs, or NVLink, per @lia_oc (October 22, 2026). Anything past 74 SMs on the 10 GB card goes through the separate RIR fuse burn, which is permanent.
Link width is a separate matter: without the PCIe x16 capacitor mod the Gen2 link runs at x4. P2P is not on cmpunlocker master; it comes from forks.
ECC: enabled, correction contested
ECC being enabled is Proven: @68x0 confirmed it live on master on October 4, 2026, and @mistyashhole ran v0.5 with ECC for about 25 hours with a clean memtest and zero ECC errors (October 10, 2026). Whether it corrects VRAM bit errors is Contested, meaning members disagree and the ECC page shows both sides. On a card with thousands of errors per memtest run, enabling ECC corrected none of them, per @zombielurker (October 5, 2026).
Upgrading and known issues
Upgrading needs no uninstall: rerun the installer as root over the old install, then cold reboot, per @phoenixone7 (October 6, 2026). A kernel upgrade leaves the card on the stock driver until you rerun the installer.
One user reported decode dropping from 60 to 80 tok/s to 35 tok/s on Qwen 3.8 27B after moving from 0.4 to 0.5, per @mowgli8628 (October 18, 2026). The cause is unconfirmed.
v0.5 is likely the last feature release, because the maintainer said little more can be had through software, per @lia_oc and @reala7096 (October 22, 2026). On October 26, 2026 @_j3ns said vGPU was close to working; the branch is public but not in master, per @arabus_74064 (October 27, 2026).