Unified Windows 10/11 (UEFI) & Linux tool + RTX 4070 Ti (AD104) validation report (RTX 40 / 30 / 20 series)

#5
by MindAllik - opened

Hello everyone,

First of all, thank you to TOTORONG for publishing the AD102 4-PLM payload and the Windows UEFI framework here, as well as the upstream researchers (Zagitta, elias-gaeros, JustAResearcher, KohakuBlueleaf, and impacifist).

Building on these foundations, I developed an open-source (GPL-2.0-only) cross-platform tool to automate the entire build and verification workflow so users don't have to manually edit local_profile.h or rely on a single pinned VBIOS ROM hash:
https://github.com/MindAllik/GPU-Tensor-Unlock-Universal

What we implemented and validated:

Hardware Validation on GeForce RTX 4070 Ti (AD104, 10de:2782): Tested on Windows 11 x64 UEFI (Palit 1569:f294, VBIOS 95.04.31.00.36, 300 W power limit). Register-resident FP16 -> FP32 jumped from 86.4 TF/s (Stock) to 172.8 TF/s (Unlocked, 2.00x), and BF16 -> FP32 reached 172.7 TF/s (2.00x), matching the card's unthrottled 172.8 TF/s FP16 -> FP16 rate.

Automatic Board Profiling (Windows_Launch_UI.bat & Linux_Launch_UI.sh): Automatically reads your card's PCI Device ID, Subsystem ID, Bus, actual VRAM window, and Windows EFI System Partition GUID, locks the PCI ID -> Silicon Die -> PMC_BOOT_0 profile, and compiles a board-matched BOOTX64.EFI onto a FAT32 USB drive.

Live Pre-Flight VBIOS Check (PromCheck): Instead of checking against a single hardcoded ROM SHA-256, BOOTX64.EFI reads the live VBIOS PROM (BAR0 + 0x300000) and runs a dry-run FRTS patch in RAM before disconnecting the UEFI display driver. If a card's VBIOS layout is incompatible, it aborts safely without leaving you on a black screen.

Multi-Generation Matrix & Built-In PTX Benchmark: Covers desktop RTX 40-series (Ada), RTX 30-series (Ampere), and RTX 20-series (Turing) profiles, and includes a 7-kernel PTX JIT benchmark (tensor_benchmark.py with randomized non-zero inputs) plus one-click NVML power limit control.

Since many desktop AD102, AD103, AD104, AD106, GA10x, and TU10x profiles in our matrix are currently marked as experimental_simulated (verified on our 106-scenario ASan simulator), we would really appreciate your feedback and hardware logs (EFI\UNLOCK\REPORT.TXT and benchmark results) - whether the unlock succeeds on your board or aborts during the pre-flight check.

Thanks for your work. It's really convinient than mine :)

Sign up or log in to comment