Русский · P2P details · Power governor
Open NVIDIA driver patches for modified CMP 50HX 20 GB / TU102 (SM75) boards. Current validated configuration: Ubuntu 24.04, kernel 6.8.0-149-generic, NVIDIA open kernel modules and userspace 610.43.03.
Checked on the four-card host on 9 November 2026:
| Item | Result |
|---|---|
| GPU memory | 4 × 20480 MiB |
| BAR1 aperture | 32768 MiB per GPU (32 GiB; this is an aperture, not additional VRAM) |
| PCIe | All four links at Gen2 x16 |
| CUDA peer capability | All 12 directed GPU pairs report access |
| Gen2 boot service | Completed successfully on kernel 6.8.0-149 |
| Idle governor | Enabled; P8 in idle, P0 under a real CUDA workload |
The current P2P patch path uses a 32 GiB BAR1 aperture and DMA32-safe bootstrap allocations. Firmware must reserve enough PCI MMIO space. A successful peer capability query alone is not a bandwidth or NCCL correctness benchmark. Historical measurements and hardware-specific limitations remain in Git history.
git clone https://github.com/rjohny55/cmp50-unlock.git
cd cmp50-unlock
sudo ./install-all.sh --install-deps --allow-experimental-gen2 --p2pThis selects the complete P2P/Gen2 patch set, stages matching kernel modules, installs the official matching userspace when needed, configures BAR1 selector 9, updates initramfs and installs the governor. It creates rollback backups and does not unload the active driver or reboot. Finish workloads and reboot manually when ready, then run:
sudo ./verify-live.sh
python3 ./verify/cuda-smoke.py
systemctl status cmp50-gen2-second-pass cmp-idle-governorTo add only the power governor to an already working driver:
sudo ./install-governor.sh --start-nowThe governor uses the panel's /data/cmp50-power control directory, preserves
existing settings, tracks driver activity history and restores automatic fan
control on load or shutdown. Its shipped binary targets Ubuntu/Debian glibc.
Use --ai-manager-data-dir /absolute/path for a different panel data mount.
CMP50_PATCH_STAGE=p2p bash ./build.sh builds artifacts for the running kernel
without installing them. The build downloads the checksum-pinned NVIDIA
610.43.03 source and applies the ordered patches. Matching kernel headers,
compiler and firmware support are required. Unknown Gen2 kernels, unsupported
PCI/subsystem identities and unsigned Secure Boot installations are refused.
This is a validated kernel target, not a promise that every newer kernel works.
Restore the printed backup with sudo ./rollback.sh /var/lib/cmp50-unlock/backups/TIMESTAMP.
Keep recovery access when changing the kernel driver or PCIe configuration.
The patched paths are board-gated; do not remove PCI identity checks. The RT-core
reporting patch does not enable physical RT execution.
Based on xrip/cmp50hx-unlock, NVIDIA's open GPU kernel modules and credited P2P work. Original research and notices are retained. See original technical material and patch provenance. Older measurements in those documents are historical, not the current installation instructions.