Skip to content

Add SBUF-resident dispatch to amortize XLA overhead for small ops #40

Description

@scttfrdmn

Benchmarks show small-op dispatch overhead dominates: even a 64×32 mask op is 96× slower on NKI than CPU. Investigate keeping intermediate tensors in SBUF across multiple kernel invocations (fused ops, persistent buffers) so the host doesn't have to round-trip XLA tensors per call.

Metadata

Metadata

Assignees

No one assigned

    Labels

    nkiNKI kernel development (Trainium hardware)

    Type

    No type

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions