liveBook · Manning

Ridge point

As explained in GPU Programming with Triton

A ridge point is the arithmetic-intensity threshold separating memory-bound and compute-bound performance on a roofline model.

On a roofline model, the ridge point is the arithmetic-intensity threshold separating memory-bound from compute-bound performance. Compare a workload’s FLOPs per byte with this threshold to identify its regime. section 3.3.1

SwiGLU elementwise operations

An arithmetic intensity of about 32 FLOPs per byte falls below the H100 Tensor Core ridge point of roughly 590 FLOPs per byte, placing the elementwise work in the memory-bound regime. section 3.3.1

How GPU Programming with Triton covers it

How deep it goes
Defines the threshold through arithmetic intensity and gives a numerical comparison, without deriving the H100 ridge point.
Examples
Compares SwiGLU elementwise intensity of about 32 FLOPs per byte with the H100 Tensor Core ridge point of roughly 590.
What you can do after
Evaluate whether a workload is memory-bound by comparing its arithmetic intensity with the ridge point.
How it connects
Connects arithmetic intensity and roofline analysis with memory-bound and compute-bound performance regimes.

Also explained in

Compare how 2 books explain ridge point →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of GPU Programming with Triton. The section links point to the parts of the book it draws on.