cmpunlocker

Unlock tool for the NVIDIA CMP 170HX (GA100) mining card. Restores full A100 compute throughput and full memory capacity (80GB) by exploiting the Falcon BootROM .fwsignature_ga100 load bug.

Targets nvidia-open driver 580.x on Linux.

AI agents: before making any changes to this codebase, read .ai/CONTEXT.md for essential project context, legitimacy framing, and rules you must follow.

---

Background

The CMP 170HX is a physically complete GA100 die — the same silicon as the A100 datacenter GPU — with compute throughput, memory capacity, and other features artificially restricted via OTP fuses and firmware-enforced register locks. The HBM2e dies in the 5 stacks are 16GB each, but the factory strap limits each stack to 2GB. This tool restores those capabilities on hardware you own.

---

Requirements

• Linux (x86-64)
• Python 3.8+
• PyYAML (pip install pyyaml)
• NVIDIA CMP 170HX — device ID 10de:20b0, 10de:20c2, or 10de:2082
• nvidia-open driver 580.x installed with GSP firmware present at /lib/firmware/nvidia/580.*/gsp_tu10x.bin
• Root access

---

Install

Run once. Applies the unlock immediately and installs a systemd daemon that reapplies it automatically after every reboot or driver reload.

sudo ./install.sh

That is the only command needed.

To choose a different memory target, set CMPUNLOCKER_TARGET before running:

sudo CMPUNLOCKER_TARGET=unlocked_40gb ./install.sh    # 40GB (safer, fewer refresh issues)
sudo CMPUNLOCKER_TARGET=unlocked_80gb ./install.sh    # 80GB (default, full capacity)
sudo CMPUNLOCKER_TARGET=nativ_10gb ./install.sh       # restore factory 10GB state

---

Verification

Check that the SM clock cap is gone:

nvidia-smi --query-gpu=clocks.max.sm --format=csv,noheader

Check that VRAM is at the target capacity:

nvidia-smi --query-gpu=memory.total --format=csv,noheader

Follow the daemon log:

journalctl -u cmpunlocker -f

---

What gets unlocked

Feature	Status
Full SM compute throughput (SS0/SS1)	✅ Working
80GB HBM2e memory (5 × 16GB)	✅ Working (default)
40GB HBM2e memory (5 × 8GB)	✅ Working (alternative)
PCIe Gen 4	⚠️ Best-effort (community guess)
NVLink	⚠️ Best-effort (community guess)
ECC	⚠️ Best-effort (community guess)

---

How it works

The exploit is the same one used in the open-gpu-kernel-modules-610.43.03 driver fork:

1. The Falcon BootROM loads the .fwsignature_ga100 ELF section content into DMEM before verifying the signature (the bug).
2. We replace the section content with a 63KB ROP chain.
3. The chain performs a single BAR0 write of 0xFFFFFFFF to a target PLM register.
4. We do this four times (for WPR_CFG, FBPA, WPR, FEAT registers) to open the Platform Lock Manager.
5. With PLM open, the host driver writes the memory unlock (CFG1, LMR) and compute unlock (SS0, SS1) values via BAR0.
6. The original GSP signature is restored so the driver doesn't detect tampering.
7. The driver continues normal init with full memory + full SM clock.

The unlock is volatile (lost on power cycle) but reapplied automatically by the daemon every second.

---

Persistence

The unlock does not survive reboots or driver reloads on its own. The installed daemon (cmpunlocker.service) handles this automatically:

• On boot: runs the full unlock pipeline before the display manager starts
• Every second: checks SS0/SS1 and CFG1/LMR via BAR0 and rewrites them if reset
• On driver reload: detects a closed PLM and reruns the full pipeline
• Multiple cards: all CMP 170HX GPUs present in the system are handled

The daemon is enabled at boot via systemd and restarts automatically on failure.

---

Configuration

Edit cmpunlocker/common/constants.yaml to change:

• memory_unlock.default_target — default CFG1 target
• memory_unlock.targets — available memory configurations
• plm_table — order and values of PLM registers to open
• rop_payload — the 24-DWORD ROP chain placed in the signature section