Turing-unlock

Board-specific research into a volatile, pre-OS tensor-compute unlock on an ASUS RTX 2080 Ti (TU102, SM75). One physical v4 run succeeded on 2026-10-22. This repository publishes the loader source, host-side tests, benchmark code and measured results. It is not a general-purpose installer or VBIOS mod.

The NVIDIA RM GET controls changed FMLA32 and IMLA4 from half to full rate. IMLA4 has not received a dedicated integer-throughput A/B test. No extra SMs, CUDA cores, Tensor Cores or VRAM were enabled. Do not interpret these results as a 2× improvement in all compute, games, AI workloads or mining.

An additional nonzero cuBLAS validation checked 295,035 outputs, all exactly matching an independent CPU reference. The 8192 case checked 128 sampled outputs, not the entire matrix. These are bounded tests, not an endurance or exhaustive numerical-stability certification.

Pearl mining with the existing PeakMiner 2.19.1 showed no meaningful gain: 85.1 TH/s after warm-up at a 190 W limit versus the user's approximate 86 TH/s historical reference. See validation notes, the comparison JSON, and the raw benchmark JSONL files under results.

• Tested unlock hardware: ASUS RTX 2080 Ti Rev. A, 10de:1e07, subsystem 1043:866a, VBIOS 90.02.17.00.47, 11 GiB, 68 SMs.
• Tested motherboard: Alienware Area-51 R2, BIOS A14, Xeon E5-1660 v3.
• Tested Linux driver: unmodified NVIDIA open driver 580.178.04.
• The successful run used Above 4G Decoding off, CSM on, and a tightly guarded direct-physical high-RAM DMA path. This is not evidence that the same path is safe on other firmware or IOMMU configurations.
• No GPU VBIOS or motherboard BIOS was flashed by this implementation.
• The unlock is volatile. A subsequent normal restart restored half-rate FMLA32/IMLA4. Automatic reapplication is not implemented or enabled.
• An RTX 5060 Ti was reinstalled and both cards resumed normal mining. The unlock has not yet been validated with both cards installed. This TU102 firmware/payload must not be applied to the SM120 card.

The source is deliberately tied to the observed PCI topology and firmware state. It checks the device at 0000:02:00.0, the exact PCI/subsystem IDs, BAR0, PCI command, parent bridge, firmware descriptors/bytes, engine state, and a finite set of observed selector tuples before executing a payload. It does not blindly choose NVIDIA GPU index 0.

Unknown pre-write states refuse the attempt. A failed, possibly partial GPU mutation holds the machine in UEFI with DMA backing retained: it must not continue into the OS; a full power cycle is required. These safeguards reduce specific risks but cannot guarantee safety or recovery on untested hardware.

The public source has an unconfigured ESP GUID, and the GRUB example has a placeholder filesystem UUID. No prebuilt EFI or automatic installation script is published. Do not remove identity/DMA guards just to make a build run. Do not run a GPU reset/unbind against the active Linux display card.

The existing one-shot GRUB entry is not suitable as a permanent default: it returns through grub.cfg, which can select it again. See persistence and dual-GPU work for the outstanding design and checks. Nothing in this repository installs a boot hook automatically.

• tools/tu102_experiment.c, tu102_experiment_core.h: v4 EFI wrapper and injected-I/O loader core.
• tools/tu102_{preflight,prom_capture,dma_services,high_dma,direct_dma}.*: logging, firmware identity and board-specific DMA guards.
• tools/test_*.c: simulated protocol/register failure-path tests; no real GPU access when running the host tests.
• tools/tensor_baseline.cu, tensor_validate.cu: SM75 throughput and independent CPU-reference numerical checks.
• tools/tu102_port_profile.py, analyze_prom_capture.py: offline ROM/profile analysis. These do not flash a card or execute firmware.
• docs/BUILD.md: dependencies, private inputs, and build/test boundaries.
• results/: measured outputs and explicit limitations.

Read BUILD.md first. The EFI/core tests require exact external firmware inputs that are intentionally not redistributed. Firmware-free startup/logging simulations can be run on x86-64 Linux with GCC and GNU-EFI:

bash tools/test_preflight.sh

CUDA 13 benchmark builds use an explicit toolkit directory:

CUDA_ROOT=/path/to/cuda-13 bash tools/build_baseline.sh
CUDA_ROOT=/path/to/cuda-13 bash tools/build_tensor_validate.sh

Building is not running. Do not run these GPU benchmarks alongside mining or other GPU jobs. They intentionally target BDF 02:00.0 and require SM75/68 SMs.

The work builds on the CMP unlock research, particularly xrip/cmp50hx-unlock, PZH1gdmu/CMP40HX-Unlock, Jon Pry's V67 canary research, and NVIDIA's open driver interfaces. See NOTICE.md for exact provenance, license notices and changes made for publication. This adaptation is distributed under copyleft-next 0.3.1, with upstream MIT notices retained.

No credentials, mining wallet, raw machine inventory, proprietary firmware, or host-specific boot installer is included.