FLOPSbased

FLOPSbased is a term used to describe a performance evaluation approach that uses floating-point operations per second (FLOPS) as the primary metric to compare computing systems and workloads. It emphasizes measuring the computational throughput of FP-bound tasks and treats FLOPS as the central unit of performance.

In FLOPSbased assessments, benchmarks consist of computational kernels or applications with known counts of floating-point operations. The score is computed as the total FLOPs performed divided by execution time, yielding a FLOPS-based score (often expressed in GFLOPS or TFLOPS). To enable cross-system comparisons, benchmarks may target FP32 or FP64 precision, use standardized input sets, and account for warmup and statistical reproducibility.

This approach is commonly used in high-performance computing and accelerator evaluation, especially for workloads where floating-point computation dominates runtime. It provides a straightforward, hardware-agnostic view of raw compute capability, and can be applied across CPUs, GPUs, and specialized accelerators when complemented with appropriate normalization.

Criticism and limitations include the fact that FLOPS-based metrics can be distorted by memory bandwidth and latency, compiler efficiency, and algorithmic choices. Real-world performance often depends on data movement and software efficiency, so FLOPS-based scores should be interpreted alongside memory-bound metrics, energy use, and application-specific benchmarks.

See also: FLOPS, LINPACK, HPL, SPEC HPC, benchmarking.