Unlock tool for the NVIDIA CMP 170HX (GA100) mining card. Restores full SM compute throughput and unlocked HBM2e memory geometry that are restricted in firmware/OTP configuration.
Join our Discord community for support and discussions.
Below are memory and performance results after applying the unlock:
Performance Benchmarks (OpenCL-Benchmark)
- Linux (x86-64)
- Root access
- NVIDIA CMP 170HX
- nvidia-open 610.43.0x already installed (libs + firmware)
- Kernel headers matching the running kernel (
linux-headers-$(uname -r)/kernel-devel) - Secure Boot disabled (patched modules are unsigned)
- Network access on first install (downloads matching stock
open-gpu-kernel-modulessources) - Python 3 (used at build time to select 8GB/10GB geometry)
To install cmpunlocker, run the following command:
sudo ./install.shTo force a certain memory profile, use the --profile option:
sudo ./install.sh --profile=8gb # 8GB card → 64GB unlock
sudo ./install.sh --profile=10gb # 10GB card → 40GB unlockThen perform a cold reboot (full power off, then boot).
| Feature | Status |
|---|---|
| Full SM compute throughput (SS0/SS1) | Working ✓ |
| Memory geometry (64GB on 8GB cards, 40GB on 10GB cards) | Working ✓ |
| PCIe Gen 2 speeds | Working ✓ |
| Full BAR1 Size (64GB) | Working ✓ |
| JTAG (Host2Jtag register access) | Working ✓ |
| Persistence across reboot (patched modules) | Working ✓ |
| Full-VRAM stability (unbacked top HBM region excluded) | Working ✓ |
To uninstall cmpunlocker, run the following command:
sudo ./uninstall.sh --yesThen perform a cold reboot (full power off, then boot).
Having issues? Need help? Join our Discord community to discuss with other users and get support.
This fork carries an extra fix on top of driver/patches/late-pma.patch: the unbacked top HBM sliver of the unlocked geometry — the last ~150 MiB with no real memory behind it — is excluded from the PMA (the late-PMA region limit is capped at 62 GiB). Upstream exposes that region to the allocator, so filling VRAM to the very top (e.g. a large matmul or a full KV cache) crashes with Xid 31 ... FAULT_INFO_TYPE_REGION_VIOLATION at a bogus address. With the clamp the region is simply dropped, leaving ~63.4 GiB of good, fully usable VRAM; allocating past it returns a clean out-of-memory error instead of a hard fault. (Clamp originally by tlswotj, verified on 3× CMP 170HX.)