cmpunlocker — CMP 90HX Compute Unlock

Full SM compute throughput restored on the NVIDIA CMP 90HX (GA102, PCI ID 10de:220d, 10 GB GDDR6X). Prefill speed went from 224 t/s → 1770 t/s (~7.9×) on a 12B Q4_0 model, with all 9 SM issue-rate fields verified full.

Primary stack: rejoin15 on NVIDIA Open 610.43.03 — V67 exploit inlined into driver init (no systemd service, survives reboots). Alternative: bendy2's V67 on 580.159.03 with systemd service.

TL;DR

What	Before	After
pp512 (12B Q4_0, prefill)	224.10 t/s	1824.15 t/s (+713%)
tg16 (decode)	throttled	55.96 t/s
SM issue-rate fields (9)	throttled (05/01/05 pattern)	all 9 full
PLM (FEAT_OVR)	locked (`0xffffff8f`)	open (`0xffffffff`)
Perplexity	202.67 (throttled, sane)	53.05 (sane)
PCIe	Gen1 x16	Gen1 x16 (hardware link cap, see Roadmap)

Verification: check.sh reports DP=full FFMA=full FMLA16=full FMLA32=full IMLA0..4=full. dmesg shows CMP90HX: V67 attempt=0 status=0x65 PLM=0xffffffff and compute selectors enabled PLM=0xffffffff SS0=0x88888888 SS1=0x00000008.

Background

The CMP 90HX is a physically complete GA102 die (same silicon as RTX 3080, sm_86, 50 SMs) with compute throttled via eFuse-set SM issue-rate modifiers. VBIOS mods cannot change eFuse defaults. The throttle is bypassed by writing the FEAT_OVR override registers (SS0=0x88888888, SS1=0x00000008) — but those registers are only writable after opening the PLM (Privilege Level Masks), which requires arbitrary code execution on the GSP Falcon coprocessor.

The vulnerability (originally disclosed by Jon Pry, "A Canary in the Crypto Mine"): the shipping GSP debug booter is encrypted with a trivial key and contains a DMA-driven buffer overflow that runs arbitrary code after signature verification, including a stack-canary bypass.

Why 610.43.03 now works (rejoin trick): earlier 610.x attempts failed because the V67 payload was injected pre-GFW — the boot-time FWSEC/WPR-meta checks rejected it. The rejoin14/15 patches move the injection to after GFW_BOOT OK: swap signature memdesc → run V67 chain (PLM opens) → driver writes SS0/SS1 → official FLR + nv_start_device retry. The 580.159.03 stack (bendy2 + systemd service) remains a valid alternative.

The Working Stack

Option A: rejoin15 on 610.43.03 (recommended)

V67 exploit inlined into driver init — no systemd service, survives reboots.

Component	What	Where
Driver	NVIDIA Open `610.43.03` + rejoin14/15 patches	kernel modules, `updates/cmpunlocker/`
Userspace	`610.43.03` (libcuda, libnvidia-ml, nvidia-smi)	from `.run` installer
GSP firmware	`gsp_ga10x.bin`	`/lib/firmware/nvidia/610.43.03/`

Install: sudo ./install.sh --profile=cmp90-rejoin15

Option B: bendy2 on 580.159.03 (legacy)

V67 via systemd service — re-applies at every boot (~2 min).

Component	What	Where
Runtime driver	stock NVIDIA Open `580.159.03` (unmodified)	kernel modules, `updates/cmpunlocker/`
Bootstrap module	`580.159.03` + bendy2's direct-compute patch (V67)	`updates/cmp90hx-persistent/nvidia.ko.bootstrap`
Service	`cmp90hx-persistent.service` — at every boot: load bootstrap → V67 → PLM open → write SS0/SS1 → 2× PCIe bus reset → reload stock driver	systemd (enabled)
Userspace	`580.159.03` (libcuda, libnvidia-ml, nvidia-smi)	from `NVIDIA-Linux-x86_64-580.159.03.run`
Kernel compat	6.12.95 build fixes for 580.159.03	`tools/fix-580-kernel612.sh`

One-Click Installer

For most users, everything is automated — download this repository (Code → Download ZIP), extract, and run:

cd cmpunlocker-master
# Recommended: rejoin15 on 610.43.03 (V67 inlined into driver, no service)
sudo ./install.sh --profile=cmp90-rejoin15

# Alternative: bendy2 on 580.159.03 (V67 via systemd service)
sudo ./install-unlock.sh

rejoin15 builds 610.43.03 + rejoin14/15 patches from source, installs to /lib/modules/$(uname -r)/updates/cmpunlocker/. Requires GSP firmware (/lib/firmware/nvidia/610.43.03/gsp_ga10x.bin) from the 610.43.03 .run installer.

bendy2 installs stock 580.159.03 + systemd service that re-applies V67 at every boot (~2 min wait after reboot).

Then: sudo reboot → verify with check.sh (9/9 fields "full"). The installer backs up your previous driver modules automatically.

Requirements: CMP 90HX (10de:220d/1555), x86_64 Linux, kernel ≥ 6.1, Secure Boot off, internet on first run.

Reproducing (for other researchers)

Prerequisites:

CMP 90HX with PCI ID 10de:220d / 10de:1555 (check: lspci -nn)
x86_64 Linux, kernel headers (/lib/modules/$(uname -r)/build), Secure Boot off

Method A: rejoin15 on 610.43.03 (recommended)

# 1. Install stock 610.43.03 (userspace + GSP firmware)
wget https://download.nvidia.com/XFree86/Linux-x86_64/610.43.03/NVIDIA-Linux-x86_64-610.43.03.run
chmod +x NVIDIA-Linux-x86_64-610.43.03.run
sudo ./NVIDIA-Linux-x86_64-610.43.03.run --silent --no-kernel-modules
# GSP firmware: /lib/firmware/nvidia/610.43.03/gsp_ga10x.bin

# 2. Build & install patched modules
sudo ./install.sh --profile=cmp90-rejoin15
sudo reboot

# 3. Verify
./check.sh   # 9/9 issue-rate fields full
nvidia-smi   # 610.43.03
dmesg | grep CMP90   # PLM=0xffffffff, SS0=0x88888888

Method B: bendy2 on 580.159.03 (legacy)

# 1. Build & install stock 580.159.03 for kernel 6.12.x
git clone https://github.com/NVIDIA/open-gpu-kernel-modules --branch 580.159.03
# apply kernel-6.12 compat fixes: see tools/fix-580-kernel612.sh
make -j$(nproc) modules KERNEL_UNAME=$(uname -r)
cp kernel-open/nvidia.ko kernel-open/nvidia-uvm.ko /lib/modules/$(uname -r)/updates/cmpunlocker/
depmod -a
# userspace: NVIDIA-Linux-x86_64-580.159.03.run --silent --no-kernel-modules
# reboot, confirm: nvidia-smi → 580.159.03

# 2. Install bendy2's persistent service
git clone https://github.com/bendy2/cmp90hx
cd cmp90hx
sudo ./install.sh
sudo reboot   # wait ~2 min for service

# 3. Verify
systemctl status cmp90hx-persistent.service     # active (exited) after ~2 min
./check.sh   # 9/9 issue-rate fields full

Rollback: sudo ./remove.sh --yes && sudo reboot

Roadmap / Remaining Work

PCIe Gen3 — the card's link is hardware-capped at Gen1 x16 (LnkCap2: 2.5GT/s). The cap is likely fuse/strap-set like the compute throttle; the now-working PLM-open path may override it via PCIE_FUSE (0x00823810), LINK_CONTROL (0x0008c000), LINK_SPEED (0x0008c040). Next experiment: extend the compute-override payload with the PCIe override. Expected: ~4× PCIe bandwidth.
Memory upgrade 10 GB → 20 GB VRAM — proposal: replace the 10× 8 Gbit GDDR6X modules with 16 Gbit (2 GB) modules (same 320-bit bus, same chip count) and unlock the 20 GB geometry via FBPA/CFG1/LMR-style overrides (the CMP 170HX 8→64 GB mechanism). Open questions: 16 Gbit GDDR6X availability, training tables, memory-controller support.
Graphics (PGRAPH2) — 3D/graphics remain gated (GSP-RM skips graphics init). Separate problem, likely needs GSP-RM firmware work.

History (what didn't work, so you don't repeat it)

Path	Result
VBIOS mods (SM issue-rate in eFuse)	❌ eFuse is hardware-locked; VBIOS cannot change it
GSP Falcon attack surface (v3–v28, 25 driver iterations)	❌ all blocked by PKC signatures / hardware firewall / PLM lock
V67 payload on 610.x pre-GFW (early attempts)	❌ FWSEC/WPR-meta checks reject pre-GFW injection
HFMA2 CUDA-core GEMM in llama.cpp (Tier 3a)	⚠️ +69% but FP16 accumulator overflow → wrong tokens; superseded by driver unlock (see `research/PREFILL_ROOT_CAUSE.md`)
V67 payload on 610.43.03 via rejoin15	✅ PLM opens (post-GFW injection)
V67 payload on stock 580.159.03	✅ PLM opens on attempt 0

Credits

bendy2 — V67 exploit + direct-compute patch for 580.159.03 (github.com/bendy2/cmp90hx) — the key that opened PLM
Jon Pry (Zenodo) — "A Canary in the Crypto Mine" (debug-booter overflow disclosure)
d3dx9 — Python Falcon emulator & ROP chain
loss-and-quick — 610.57.04 port, gadget analysis, tools, English translations (merged PR #2)
Rhonstin — CMP 90HX patches
amoghmunikote — original cmpunlocker for CMP 170HX (GA100)

Repository Layout

STATUS.md — full research status (registers, exploit chain, benchmark data)
docs/ — exploit writeup, bendy2 analysis, gadget analysis (580.x vs 610.x), testing guide
driver/ — build scripts + patches (610.43.03 era and 610.57.04 port)
tools/ — analysis tools (lw_catalog_610.py, compare_op32.py, extract_ucode.py, fix-580-kernel612.sh)
research/ — Falcon attack surface iterations, SM issue-rate comparison

License

Same as original cmpunlocker

Name		Name	Last commit message	Last commit date
.github		.github
analysis		analysis
common		common
docs		docs
driver		driver
research		research
tools		tools
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
STATUS.md		STATUS.md
build-cmp90-610.sh		build-cmp90-610.sh
build-cmp90.sh		build-cmp90.sh
cmp90_rpc_log.txt		cmp90_rpc_log.txt
install-unlock.sh		install-unlock.sh
install.sh		install.sh
kernel61-fix.sh		kernel61-fix.sh
remove.sh		remove.sh

Navigation Menu

Repository files navigation

cmpunlocker — CMP 90HX Compute Unlock

TL;DR

Background

The Working Stack

Option A: rejoin15 on 610.43.03 (recommended)

Option B: bendy2 on 580.159.03 (legacy)

One-Click Installer

Reproducing (for other researchers)

Method A: rejoin15 on 610.43.03 (recommended)

Method B: bendy2 on 580.159.03 (legacy)

Roadmap / Remaining Work

History (what didn't work, so you don't repeat it)

Credits

Repository Layout

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

cmpunlocker — CMP 90HX Compute Unlock

TL;DR

Background

The Working Stack

Option A: rejoin15 on 610.43.03 (recommended)

Option B: bendy2 on 580.159.03 (legacy)

One-Click Installer

Reproducing (for other researchers)

Method A: rejoin15 on 610.43.03 (recommended)

Method B: bendy2 on 580.159.03 (legacy)

Roadmap / Remaining Work

History (what didn't work, so you don't repeat it)

Credits

Repository Layout

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages