FLOPrate

FLOPrate is a metric used to quantify the sustained throughput of floating-point operations on a computing system, expressed in floating-point operations per second (FLOPS). Unlike peak FLOPS, FLOPrate reflects actual performance under a defined workload and software stack, including compiler optimizations and library calls.

Calculation of FLOPrate relies on a benchmark or workload that specifies a known total number of floating-point operations. FLOPrate is then computed as the total number of operations performed divided by the elapsed wall-clock time. Operation counting follows standard conventions: a multiply or add counts as one FLOP; a fused multiply-add can be counted as either two FLOPs or one, depending on the adopted convention.

Measurement and normalization practices vary. To enable cross-system comparisons, FLOPrate is often reported for a representative kernel mix (for example, dense linear algebra, FFTs) or a full benchmark suite. Reports may distinguish between single-precision and double-precision FLOPrate, and may adopt different counting schemes for fused operations.

Interpretation and limitations are important. FLOPrate is sensitive to memory bandwidth, cache hierarchy, and compiler/runtime optimizations, so it should be viewed as sustained performance for the tested workload rather than a universal capability of the hardware.

Applications include benchmarking hardware vendors, evaluating system configurations for HPC centers, and guiding software optimization. FLOPrate complements peak FLOPS and traditional benchmarks, providing a practical measure of real-world floating-point throughput under specific workloads.

See also: FLOPS, GFLOPS, benchmark, high-performance computing, performance metric.