AI inference glossary

Software

Asynchronous copy

Also known as async copy, cp.async, LDGSTS

In plain English

A data transfer that can proceed while independent computation runs, with synchronization before consumers use the result.

Technical definition

An asynchronous copy decouples initiating a transfer from waiting for its completion, allowing independent operations to execute in between.

Engineering details

The Blackwell port focuses on the cp.async global-to-shared path, which avoids staging data through ordinary registers. Thread count, load width, pipeline stages, and outstanding bytes determine resource demand. Barriers or completion mechanisms remain necessary before the destination can be consumed safely.

Why it matters

Asynchronous submission does not guarantee overlap or higher bandwidth. A dependent consumer, insufficient buffering, or memory-pipeline congestion can still serialize execution or make additional stages counterproductive.

How to read it in InferenceX

Compare LDGSTS and TMA with equivalent data and clearly stated configurations. The article’s measured latency and throughput are distinct quantities and cannot be exchanged as a universal ranking of copy mechanisms.

See the concept in real benchmarks

All articles