FLOPs

FLOPS stands for floating-point operations per second and is a unit of computing performance that measures how many floating-point arithmetic operations a processor can perform each second. A single floating-point operation is a FLOP, and FLOPS refers to the rate of performing such operations. The exact counting of what constitutes a FLOP can vary; an addition or multiplication is typically one FLOP, while a fused multiply-add (FMA) may be counted as two FLOPs or as one, depending on the convention used.

FLOPS are commonly expressed with prefixes such as MFLOPS, GFLOPS, TFLOPS, and PFLOPS to reflect different scales of performance. Theoretical peak FLOPS are estimated from a processor’s clock rate and its parallel execution resources, such as vector width and the number of execution units. For example, a core that can perform multiple floating-point operations per cycle across a wide vector unit at a given clock rate yields a higher peak FLOPS figure.

In practice, achieved FLOPS are typically lower than the theoretical peak due to memory bandwidth, latency, and algorithmic efficiency. Real-world performance is often assessed with benchmarks like LINPACK or SPECfp, which measure sustained FLOPS under representative workloads rather than peak capability. FLOPS remain a standard measure of raw arithmetic capacity, especially in high-performance computing, but comparisons require specifying the precision (single vs. double), the counting convention, and whether the figure reflects peak or sustained performance.