Source-first research on volatile NVIDIA tensor-compute unlocks. Not a universal GPU unlocker.
RTX 2080 Ti Linux/initramfs and RTX 4090 Windows/UEFI unlocks, with RTX 3090 upstream deployment experience. Includes source, offline tests, firmware/ROM analysis guidance, boot workflow documentation and measured results.
- 2080 Ti: physically VRAM-modded Leadtek cards, 11 GB → 22 GB.
- 4090: AD102 USB EFI transaction, stock Windows NVIDIA driver handoff, runtime-randomized instruction benchmarks.
- 3090: four-card deployment using the upstream Ampere implementation.
Firmware, ROM dumps, signing keys and compiled kernel/EFI binaries stay local.
Risk: experimental privileged firmware execution can hang a GPU or host, prevent boot, corrupt data, or damage hardware. There is no guarantee of recovery. Have physical/BMC console and power control. Never experiment on an active workload. Editing IDs or removing guards does not validate a new card.
The 22 GB cards are physically VRAM-modded, not factory 22 GB models. Both the unlocked card and locked comparison card already had that modification. This software unlock does not add VRAM, and the repository does not provide a memory-soldering guide.
| Test | Locked | Unlocked | Comparison |
|---|---|---|---|
2080 Ti FP16→FP32 mmapeak, 32×8×16 |
59.9 TFLOPS | 111.0 TFLOPS | Same card before/after; +85.3% |
| 2080 Ti FP16→FP16 control | 119.5 TFLOPS | 118.5 TFLOPS | Same card; essentially unchanged |
| 2080 Ti best MAMF median, 125-shape sweep | 57.71 TFLOPS | 88.24 TFLOPS | Two same-model cards; +52.9%, not same-card A/B |
4090 FP16→FP32 randomized mmapeak |
178.9 TFLOPS | 331.5 TFLOPS | Same card/binary; +85.3% |
4090 BF16→FP32 randomized mmapeak |
179.0 TFLOPS | 330.9 TFLOPS | Same card/binary; +84.9% |
| 4090 FP16 MAMF Finder, best median of 125 shapes | ~160 TFLOPS | 243.9 TFLOPS (measured) | WSL2; 4096x5120x2048 |
| One 3090, FP16→FP32 8192³ PyTorch GEMM | 73.03 TFLOPS | 122.76 TFLOPS | Same card; +68.1%* |
*3090 power limits: 370 W stock / 360 W unlocked. Measurement details.
See results and limitations. A successful kernel log alone is not proof of higher throughput. BF16/TF32 are not added to Turing. This is not a claim of a 2× application/game speedup.
- Safety and scope
- Agent engineering guide: investigation, porting, failure lessons
- 2080 Ti reproduction: inputs, build, one-shot boot, verification, recovery
- 3090 experience: upstream and known-working revision
- 4090 Windows/UEFI source, results and references
- Agent-oriented Ada porting guide: inventory, ROM extraction, FWSEC parsing, profile adaptation, ownership, tests and signing
- Licensing/provenance
4090 results use our mmapeak fork. The original compile-time all-zero fragments allowed the compiler to collapse identical tile calculations while the host still counted full-tile FLOPs in inspected builds. Our fork uses runtime-randomized inputs. See the analysis and pinned revision.
# Linux, Python 3.10+, C compiler; no GPU or sudo needed
make test
# Protocol simulation with privately supplied firmware; no GPU access
make test-private ASSETS=/absolute/path/to/verified/assets
# Compile only against matching installed kernel headers
make modulemake module builds a hardware-execution-disabled module by default. A reviewed exact-profile experiment can opt into compilation with make module ENABLE_EXPERIMENT=1; runtime parameters are still required. No target installs/loads modules, edits boot configuration, flashes anything, or runs external installers.
python3 tools/inventory.py --bdf 0000:42:00.0 --output local/inventory.json
python3 tools/rom_analyze.py local/card-prom.bin --extract local/native-fwsec.bin
python3 tools/check_assets.py --assets local/assetsInventory uses OS metadata and nvidia-smi, not BAR mmap or resets. Output is private and gitignored. ROM analysis is file-only; obtaining a full dump is a separate hardware operation covered in the guide.
- Two Leadtek TU102 rev a1 cards:
10de:1e04 / 107d:1e04, VBIOS90.02.33.00.3F, physically VRAM-modded to 22 GiB, 68 SMs; kernel 6.8.0-147 and NVIDIA open 595.99.02. Both unlocked in the local sequential implementation. The public executor remains single-target. - Two-card NCCL all-reduce passed from 1 KiB to 64 MiB with NV2 topology, bidirectional peer access and NCCL direct P2P. At 64 MiB: 36.77 GB/s algorithm bandwidth.
- GIGABYTE MZ32-AR0 / AMD server, separate ASPEED display, Secure Boot disabled. Shared IOMMU group observed: no bus-wide resets.
- Turing public reproduction: one-shot entry and mandatory initramfs gate; local combined 3090+2080Ti persistent setup was separately validated. Dual-card notes.
- Ampere: external Zagitta/gpu-unlock; not vendored here.
- Ada: one RTX 4090 Windows/UEFI manual USB unlock on ASUS Pro WS W890E-SAGE SE, NVIDIA 621.01, 450 W. Signed EFI, successful trial with Secure Boot off. Source and results.
- Other board profiles and Blackwell remain research targets.
Public source derives from the tested source with a default-off compile gate and publication packaging changes. It is not byte-identical to the historical test binary. No public-source hardware rerun is claimed.
Based on haohervchb/Turing-unlock, xrip/cmp50hx-unlock, CMP40HX work, Jon Pry's V67 research, and NVIDIA driver interfaces. We do not claim discovery of the exploit. See NOTICE.md.