Performance Analysis
When the Roofline model shows a workload in the memory-bandwidth-limited region, which direction improves performance?
- AConsider only adding more arithmetic units
- BIt will naturally speed up over time without changes
- CRaise arithmetic intensity, for example by increasing data reuse, or reduce memory traffic
Show answer and explanation
Answer: C. Raise arithmetic intensity, for example by increasing data reuse, or reduce memory traffic
In the Roofline chart, workloads under the slanted line are capped by memory bandwidth, and adding arithmetic units does not raise that cap. Reusing data through blocking or operator fusion increases operations per byte (arithmetic intensity), moving the workload rightward and raising its ceiling. Using lower-precision data to cut memory traffic has a similar effect.
- A. When memory-bound, more arithmetic units do not help.
- B. Without changes, it will not get faster.
- C. Correct. Move rightward toward the flat part of the roof.