laskentaintensiteettiä

Computational intensity, or laskentaintensiteetti in Finnish, is a key metric in high‑performance computing that quantifies the ratio of arithmetic operations to memory operations executed by an application. It is usually expressed as floating‑point operations per byte of data moved between memory and processor. A high computational intensity indicates that a program performs many calculations for each memory read or write, which tends to favor CPU or GPU compute performance over memory bandwidth. Conversely, a low intensity reveals a memory‑bound workload where data transfers dominate the execution time.

The concept underlies performance modeling frameworks such as the Roofline model, which plots attainable performance against operational intensity to identify the limiting hardware resource. Designers use computational intensity to assess whether hardware upgrades (e.g., adding more cores or higher memory bandwidth) will effectively speed up a workload. In many scientific and engineering codes—such as finite element solvers, image processing, and machine learning training—the algorithms can be tuned to increase intensity by restructuring data layouts, employing blocking techniques, or fusing loops to reduce data traffic.

Measuring intensity involves profiling the code to count operations and memory accesses accurately, often using instrumentation tools or compiler analysis. Standard practice also includes normalizing the count to a standardized set of operations (e.g., multiply–add pairs) for comparability. While the metric is simple, it provides valuable insight into how software interacts with heterogeneous architectures, guiding both software optimization and hardware procurement decisions.