MATHEMATICAL GPU SYSTEMS · FROM INVARIANT TO KERNEL

#1 GPU MODE · B200 · zhongmingee Cholesky · 0.201531 ms online ↗

6 CHAPTERS · FROM FIRST PRINCIPLES TO THE FROZEN KERNEL

Cholesky factorization,
derived for the GPU

    PROVENANCE APPENDIX Open the retained route metadata Repository artifact hash, environment, and route context; online competition identity remains a separate record.

    From invariant to implementation

    ARTIFACT
    CHOLESKY
    SIZE
    1,091,590 B
    GPU
    NVIDIA GB200
    ISA
    SM100 · CC 10.0
    CUDA
    13.0
    TRITON
    3.7.x
    COMMIT
    66755e3
    CONCEPTUAL APPENDIX · NOT A GPU TRACE Open the classical blocked dependency sandbox Fixed teaching costs model causality only; they are not B200 timing, occupancy, or SM placement.

    Replay the dependency chain

    This model visualizes the classical blocked grammar derived above. Production kernels fuse some boundaries; one node is not one launch.

    VIEW A

    Tile state

    VIEW B

    Model schedule

    Generated scheduler output

    VIEW C

    Task dependency DAG

    Solid nodes are complete · edges come from depends_on

    WHAT TO NOTICE

    Measurement after derivation

    The selected algorithm’s raw GB200 rows are shown at readable size. Correctness and evidence boundaries remain part of the result; the linked GPU MODE placement is kept separate from this local capture.

    Loading checked-in benchmark rows…