GFLOPS

GFLOPS stands for giga floating-point operations per second and is a standard measure of a computer’s floating-point compute performance. It expresses how many floating-point operations a device can perform every second, typically for single-precision (32-bit) or double-precision (64-bit) arithmetic. A floating-point operation refers to basic math like addition or multiplication of real numbers.

GFLOPS can denote both peak theoretical performance and measured sustained performance. The peak value is determined by the hardware’s floating-point execution units and its clock frequency; for devices with fused multiply-add units, an FMA instruction may count as two floating-point operations, depending on counting conventions. Because different benchmarks and vendors may adopt different counting rules or precisions, reported GFLOPS figures can vary, so comparisons should specify the precision and the counting method used.

In practice, GFLOPS is widely used to compare CPUs, GPUs, and accelerators, especially for scientific computing and machine learning workloads. Values are commonly reported in GFLOPS, TFLOPS, or PFLOPS, reflecting the scale of the hardware. The general idea is that peak GFLOPS provides a rough upper bound on how fast a system can perform floating-point computations, while sustained GFLOPS give a more realistic measure of performance under real workloads.

GFLOPS has limitations as a sole metric. Real-world performance also depends on memory bandwidth, latency, algorithm efficiency, compiler optimizations, and parallelism. Therefore, GFLOPS is most informative when used alongside other metrics and within the context of specific applications and benchmarks.