AI inference glossary

Benchmark metrics

Pointer chasing

In plain English

A dependent sequence of memory loads used to probe latency while limiting the overlap available between accesses.

Technical definition

Pointer chasing follows a chain in which each loaded value determines the address of the next load, creating a serial memory dependency.

Engineering details

Because the next address is unknown until the current load returns, the test limits memory-level parallelism. Cache residency, address distribution, chain length, and concurrent traffic still affect the measured latency. Comparing latency profiles across cores can suggest shared resources or topology.

Why it matters

A latency-distance matrix is indirect evidence, not a physical floorplan. Contention and address mapping can produce apparent groups, and a dependent-load result cannot substitute for sustained bandwidth under many independent loads.

How to read it in InferenceX

The Huawei and Blackwell ports use latency probes to infer resource relationships. Preserve their distinction between observed timing, inferred logical groups, and speculative physical placement.

See the concept in real benchmarks

All articles