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.
This one was fought by an "AI fleet" — GPT-5.6 Pro, Kimi K3 and other top-tier models running as a multi-agent team: dedicated agents for reverse engineering, firmware analysis, and live hardware experiments. A week of work, 60+ controlled experiments. This is the 80GB memory story.
✅ What we achieved: 80GB geometry unlocked
nvidia-smi shows 81920 MiB. The physical 80GB of HBM is confirmed present and addressable, and the card ran long-context LLM inference with our memory-safety scheme.
🧱 But there's a wall: the 40G fold Writes above 40G (= 20 FBPA × 2G) fold back to low addresses at a +35GiB offset (256B-granular tail, 8G = HBM-die periodicity). We fully decoded the wall's signature first, ruled out page-table/driver-layer causes, and pinned it to the FBPA routing layer.
🔬 Then the hard part — exhausting every software path (each with hard evidence):
- SEC2 HS (LEVEL2) register writes — reproduced the paper-grade primitive; all 9 writable FBPA registers, single and combined writes, persistence verified → wall unmoved
- Regkey reroutes (FBFLCN disable / engine switch / PMU) → all dead
- Host-side VBIOS table patching — we actually won the authentication bet (MAC-region edits are never re-verified on upload!), but FWSEC reads the table straight from the card's ROM; it never touches the host buffer → dead end
- VBIOS flashing assessment — 8-9/10 odds of MAC (RSA-3072) rejection + brick risk → rejected
- WPR2 runtime patch (the last hope) — built a SEC2 post-auth DMA injection platform to patch GSP-RM at runtime and rewrite FWSEC's staged table copy → the GSP-RM app layer has no memory window reaching that region; four anchor attempts, all severed
🎯 Root cause (a surprise payoff from the AI team's VBIOS table RE): The real CMP-vs-A100 delta isn't at the register level — it's 19 per-partition dwords (columns 7/8) in the VBIOS devinit table, selected by the fused device-id inside the signed + encrypted FWSEC ucode running on SEC2.
Final verdict: the SKU boundary is an output of the silicon-anchored chain of trust — short of NVIDIA's signing key or fuse-burning authority, linear 80GB is not software-unlockable, period. We've nailed down the evidence at every layer, so nobody has to walk this maze again.
Full docs, 22 RE notes, complete experiment log (v1–v64): docs/FINAL_VERDICT_40G_WALL.md · docs/RESEARCH_INDEX.md
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 the compiled memory 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 unlockThe stable default for a 10de:2082 10GB card remains 40GB. An explicit
experimental profile is also available:
sudo ./install.sh --profile=10gb80 # 10GB card → experimental 80GB geometryThis compiles the coherent 80GB values into the real driver path:
CFG1=0x02779000, LMR=0x0000028B, and
fb_length=0x0000001400000000. It is never selected automatically. On a mixed
20c2+2082 system, 20c2 cards stay on 64GB while 2082 cards use the experimental
80GB target.
Production recommendation (Aug 2026 verdict): use 40GB (--profile=10gb).
The 80GB profile is fully validated (gpu_burn 300s clean, 256K-context LLM
inference working) but the 40G fold wall caps its safely usable pool at
~31–36GiB ([5G,36G) between the fold-impact zone and the phantom-reserve hole),
which is less than the clean 40GB profile — and any allocation crossing 36G
risks silent fold corruption. The 80GB profile remains as a research artifact;
see Final verdict on the 40G wall and
Experimental 80GB.
This source includes the wpr-safe-r3 fix. The former experimental late-PMA
path that converted the highest reserved FB region into allocatable memory has
been removed. GSP WPR, firmware heap, metadata, and other bRsvdRegion
carveouts remain reserved and are never passed to pmaRegisterRegion() by
cmpunlocker.
The safe allocatable amount can therefore be slightly lower than the capacity
shown by nvidia-smi; that difference is expected firmware/driver reservation,
not missing user memory. build.sh and verify.sh reject modules containing
the removed late-PMA marker.
The installer does not hot-reload the GPU driver by default. Perform a complete power-off/cold boot, then run:
sudo ./verify.sh
sudo ./tools/collect-diagnostics.shFor the first high-memory workload, run it through the monitor so both the pre-failure state and the first kernel error are preserved:
sudo ./tools/run-monitored.sh --interval=1 --output=/root/cmp-logs -- \
python3 your_workload.pyThe monitor writes a timestamped archive plus SHA-256 checksum even when the
workload exits with an error. Do not use CMPUNLOCKER_ALLOW_HOT_RELOAD=1 for
stability qualification; that developer override cannot prove stale GSP/WPR
state was cleared.
| Feature | Status |
|---|---|
| Full SM compute throughput (SS0/SS1) | Working ✓ |
| Memory geometry (64GB on 8GB cards, 40GB on 10GB cards) | Working ✓ |
| 80GB geometry on 10GB cards | Working ✓ but 40GB recommended for production — the 40G fold wall caps 80G's safe pool below the 40G profile's (see verdict below) |
| PCIe Gen 2 speeds | Working ✓ |
| PCIe Gen 3/4 | Not unlockable — blocked at fuse/PHY-cal layer (see verdict) |
| JTAG (Host2Jtag register access) | Working ✓ |
WPR/PMA reserved-memory protection (wpr-safe-r3) |
Working; old unsafe module rejected |
| Persistence across reboot (patched modules) | Working ✓ |
A week-long deep investigation pushed the 10GB card to full 80GB geometry and then hunted the "40G wall" (writes above 40G fold back with a +35GiB alias) all the way to its root cause. Full archive: docs/RESEARCH_INDEX.md; final verdict: docs/FINAL_VERDICT_40G_WALL.md.
Milestones
- Aug 6–8 — 80GB geometry unlock stabilized (81920 MiB in
nvidia-smi), phantom-reserve memory-safety scheme, full compute throughput restored (FP32 0.39 → 12.26 TFLOPS, BF16 166 TFLOPS class). - Aug 8–9 — The "32G/40G wall" discovered and independently re-confirmed; wall signature fully decoded: +35GiB fold with a 256-byte-granular tail, 8G (HBM-die) periodicity (docs/WALL_ALIAS_DECODE.md). Hole narrowed 8G→5G.
- Aug 10 — SEC2 post-authentication DMA injection platform built: runtime GSP-RM patching in WPR, LEVEL2 (HS) register writes, VBIOS-table patch hooks — all without touching any signature (docs/PLAN_SEC2_DMA_POSTPATCH.md).
- Aug 10–11 — Systematic elimination of every software-reachable layer: all 9 writable FBPA registers (single + combined writes, persistence-verified), 3 regkey reroutes, host-side VBIOS table patching, VBIOS flashing risk analysis, and the WPR2 staging-window (R1) probe. VBIOS table semantics cracked: the real CMP-vs-A100 delta is 19 per-partition dwords (columns 7/8) selected by fused device-id inside the signed/encrypted FWSEC devinit ucode (docs/re/A100_40G_80G_COLUMN_DIFF.md, docs/re/LATE_OVERRIDE_0294.md).
- Aug 11 — Final verdict: the 40G wall (and PCIe Gen3/4) is a fuse-selected, cryptographically authenticated SKU decision; short of NVIDIA's signing key or fuse authority it is not software-unlockable. Shipped production configuration: 80GB geometry + 5G phantom reserve, validated end-to-end (gpu_burn 300s, 68GB occupied, 0 errors, zero Xid; llama 256K-context inference working with KV cache placed below the fold).
Tooling left behind: BAR0 MMIO read/write probes, wall/fold scanners and the alias decoder
(tools/), the SEC2 probe generator framework (driver/apply_sec2_dma_probe.py), and 22
reverse-engineering notes (docs/re/). Firmware binaries and disassembly are
intentionally not published.
To uninstall cmpunlocker, run the following command:
sudo ./remove.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.