Search ARuntime.com

Find runtime definitions and implementation guidance

Search page titles, summaries, headings, glossary terms, use cases, and runtime-directory entries.

Enter at least two characters.

Runtime Stack

Kernels and Hardware Libraries

Tensor kernels, device libraries, collectives, communication paths, and low-level performance boundaries.

Audience: ML systems engineers; compiler and performance engineers Reading time: 1 minute Status: Architecture Last reviewed:

Kernel and hardware libraries implement the primitive operations that compilers and inference engines schedule onto devices.

Responsibilities

  • Matrix multiplication, convolution, attention, normalization, sampling, and vector operations.
  • Collective communication, topology-aware transfers, and device synchronization.
  • Precision-specific paths, layout requirements, workspace allocation, and autotuning.
  • Versioned compatibility with drivers, accelerators, and compiler-generated code.

Operational evidence

Record kernel selection, precision, fallback, launch overhead, device occupancy, memory traffic, and communication time when diagnosing performance.

Maintenance record

Found an error, outdated capability, or unclear category boundary? Submit a correction with a supporting source.