Skip to content

Your GeForce runs AI at half speed. This fixes it for 20/30/40 series cards. Up to 2× FP16 / BF16 / TF32 tensor throughput

C
2.4 KB
Raw Blame

Firmware provenance and distribution scope

gsp_ga10x.bin and gsp_tu10x.bin are extracted byte-for-byte from the NVIDIA 610.43.03 package by extract_gsp.py. The Booter and GSP boot images are decoded unchanged from the pinned 610.43.03 source archive. The extraction also preserves that package's complete usr/share/licenses/nvidia-utils/LICENSE as NVIDIA-LICENSE. It is installed with the firmware package.

The source archive's open-module licensing does not license NVIDIA firmware. The binary package contains the NVIDIA Driver License Agreement. Its section 1.1(d) addresses distribution for open source operating-system kernels with unchanged binary files and inclusion of the agreement. Additional limitations apply, including NVIDIA-platform use. Personal local packaging is the initial delivery target. Review the full pinned agreement and the separately generated payload before any public distribution; redistribution approval is not claimed.

The GA102 re-entrant payload is derived from the live 610.43.03 Booter behavior (its OS-block validation, PLM pair writes and exit path), not from the cmpunlocker template. The Stage 1 reference checkout, amoghmunikote/cmpunlocker at commit 2d10f891a58824f0591c3332c4eccf37ff36b086, is the recorded origin of the initial payload constants and PLM group. Its repository license is GPL version 2, preserved in licenses/GPL-2.0.txt. The payload is source code in this repository, separate from the unchanged NVIDIA firmware.

NVIDIA source files inspected during the loader audit carry their own MIT copyright notices. The minimum helper's metadata layout and Falcon operations follow those source contracts; licenses/NVIDIA-MIT.txt preserves the notice. The inspected EFI reference's notice is in licenses/CMP40HX-MIT.txt. The payload reference targets GA100 CMP 170HX. It does not establish GA10x firmware acceptance, standalone multiwrite, final FLR retention or stock-driver handover. Those require the Stage 1 hardware experiments.

The Linux loading geometry, register sequence and success gates are ported from elias-gaeros/gpu-unlock's RTX 4090 work. That report records one successful RTX 4090 production-booter transaction. A hardware repeat of the refactored path is still needed. AD102 uses the same per-entry signature image as GA10x. Its privileged stores are WPR_CFG, WPR and FEAT.