Why it matters

Roofline models GPU perf. Understanding shapes optimization intuition.

Advertisement

The architecture

Peak FLOPs: hardware limit.

BW * AI: BW-limited plot line.

Intersection: knee.

Roofline modelPeak FLOPshorizontal ceilingBW * AIslope up to kneeOp positionon curvePosition ops on plot; move toward compute-bound via fusion + higher AI
Roofline.
Advertisement

How it works end to end

Ceiling: peak GPU FLOPs.

Slope: BW * AI.

Ops sit on curve.

Move op right (higher AI) to hit peak.