NVIDIA build — 3 lines × 4 formats (HiveOS / SMOS / mmpOS / Linux).
New: pascal line — GTX 10-series / Tesla P100 (sm_60, e.g. GTX 1080 Ti). CUDA 12.4, driver floor R550. The PoM walk runs full-speed on Pascal; run inference on CPU: add --cpu-inference --light.
Lines & driver floor:
legacy— CUDA 12.4, driver ≥ R550 (Turing 20xx / Ampere 30xx / Ada, older drivers)modern— CUDA 12.9, native sm_120, driver ≥ R575 (Blackwell 50xx / PRO)pascal— CUDA 12.4, sm_60, driver ≥ R550 (GTX 10-series)
Pick the tarball for your OS (HiveOS/SMOS/mmpOS/generic Linux). Verify with the matching SHA256SUMS-<line>.txt. AMD and Windows are distributed separately.