FLOPS
A measure of computing performance counting floating-point operations per second.
Definition
FLOPS (floating-point operations per second) measures how many real-number arithmetic operations a system performs each second. It is the standard yardstick for scientific computing performance, scaling through mega-, giga-, tera-, peta-, and exa-FLOPS.
Roofline analysis places an application on a chart of arithmetic intensity against performance, showing whether it is limited by computation or by memory bandwidth. This diagnostic guides optimization more usefully than a single FLOPS number, by naming the actual bottleneck.
Peak FLOPS is a ceiling rarely approached in practice, because most real applications spend much of their time waiting for data rather than computing. The roofline model makes this concrete by plotting achievable performance against arithmetic intensity, revealing whether a code is limited by computation or by memory bandwidth. Naming the true bottleneck this way guides optimization far more usefully than chasing a higher peak FLOPS figure.
Caveats
- Peak FLOPS is theoretical; sustained FLOPS on real code is lower.
- Memory bandwidth often limits achievable FLOPS.
- It ignores integer work, communication, and I/O.
Why it matters
FLOPS gauges the raw arithmetic capacity available for simulation, and supercomputers are ranked by it. But most real applications are limited by data movement rather than arithmetic, so high peak FLOPS does not guarantee proportionally fast programs.
Fusion connection
The FLOPS available to Kronos sets an outer bound on the resolution and number of Hyperion simulations feasible, though in practice memory bandwidth is often the tighter constraint.