AI inference glossary

Software

CANN

In plain English

Huawei’s software stack for compiling and running AI operations on Ascend hardware.

Technical definition

CANN is the Ascend software stack that supplies programming, compilation, runtime, and operator facilities for neural-network execution.

Engineering details

A framework backend depends on more than a compiler accepting its program. Operators, runtime interfaces, drivers, and communication components must agree on versions and supported hardware. The Huawei article discusses both CANN-provided operators and kernels authored through other tools, including TileLang.

Why it matters

Comparing two frontends without identifying their underlying runtime and operator paths can attribute a performance difference to the wrong component. Installation compatibility is also distinct from numerical correctness and measured serving performance.

How to read it in InferenceX

Treat release-status statements in the port as historical. Its CANN and HDK paragraphs retain different drafting dates, so they are not a current version-selection guide.

See the concept in real benchmarks

All articles