Your GeForce runs AI at half speed. This fixes it for 20/30/40 series cards. Up to 2× FP16 / BF16 / TF32 tensor throughput
Pinned 610.43.03 firmware
make inputs downloads the two NVIDIA inputs into the cache.
make firmware runs extract_gsp.py for the two GSP ELFs and the NVIDIA
license, and extract_boot_assets.py for the signed SEC2 Booter bundles,
generic Falcon loader, and per-family GSP boot images/descriptors.
Extracted NVIDIA bytes are unchanged.
The payload is ordinary source code: the V67 call table in
kernel/gpu_unlock/recipes.c, built per write by payload.c. There is no
generator, no external dmem.bin override, and no reference image on disk.
The SEC2 Booter executes on Falcon and loads the separate GSP boot image.
GA100 uses gsp_tu10x.bin with .fwsignature_ga100; GA10x uses
gsp_ga10x.bin with .fwsignature_ga10x. The selected card determines its
GSP assets once, separately from register tables. All new hardware contexts
remain experimental; see Stage 1 evidence.
Python 3, zstd, and the pinned inputs are sufficient for extraction. See NOTICE.md for provenance and licensing.