make inputs runs fetch_artifacts.py, which downloads the NVIDIA open kernel
module source archive and the exact CachyOS nvidia-utils package pinned in
artifacts-610.43.03.json. Downloads use HTTPS and are checked against their
SHA-256 before they enter the cache; cached files are rechecked on every run.
Nothing downloaded is executed or installed.
make firmware then builds the nine files the kernel module loads from
gpu-unlock/610.43.03/:
generate_v67.pyextractsgsp_ga10x.binand the NVIDIA license from the package and generates the 64,000-bytev67-ga102-610.43.03.binandv67-ad102-610.43.03-four-plm.binpayloads.extract_boot_assets.pybuildsbooter_load_ga102.bin,gsp_boot_ga102.binandgsp_boot_desc_ga102.bin, plus their three AD102 equivalents, from the source archive.
The AD102 payload reproduces the four-write chain pinned in the
manifest at SHA-256 44c36b2d6b3424e53e18f3e55db85c833655de071ce66ce42749588a4a7dc8cb.
Its ordered targets are WPR_CFG, FBPA, WPR and FEAT_OVR_PLM; the first value is
0xfffff0ff, and the other three are 0xffffffff. It then requests mailbox0
marker 0x67b00001 and cleanup. Runtime readback checks the marker, required
permission registers and SEC2 HALT before applying selectors. The
RTX 4090 report records production execution,
FLR/stock-driver persistence and approximately doubled FP32-accumulate tensor
throughput on one card. This payload and its plaintext-derived instruction
addresses come from TOTORONG's AD103_win_unlocker.
The manifest records payload offsets separately from DMEM addresses. With the
modeled copy base 0x600, the canary at payload 0xf94c lies at DMEM 0xff4c,
and the return slot at payload 0xf950 lies at DMEM 0xff50. The AD102
FEAT_OVR_PLM address word is at payload 0xf9cc; GA102 uses 0xf960.
Both scripts check every input and output against its pinned SHA-256 before
writing anything. Python 3 and the zstd command are required for extraction. See
NOTICE.md for provenance and licensing.