TFLOP

TFLOP stands for terafloating-point operations per second. It is a unit of computational performance describing how many floating-point calculations a computer can perform in one second. One TFLOP equals 10^12 floating-point operations per second. In practice, the rate is reported for a specific precision and operation type, since single-precision (32-bit) and double-precision (64-bit) results differ and conventions for counting operations (for example, how a fused multiply-add is counted) can vary.

A floating-point operation involves arithmetic with real numbers, such as addition, multiplication, or fused multiply-add. Because FLOPs are not directly observable, TFLOPS ratings are usually given as peak theoretical performance or sustained performance on benchmarks. The terms Rpeak (theoretical peak) and Rmax (achieved sustained performance) are commonly used in the context of HPC benchmarks like LINPACK.

TFLOP ratings are often discussed in relation to precision: a device may deliver many hundreds or thousands of single-precision TFLOPS, but fewer double-precision TFLOPS. Real-world performance depends on algorithms, memory bandwidth, and software efficiency, not solely on raw arithmetic throughput.

For scale, 1 petaflop equals 1,000,000 TFLOPS. TFLOPS provide a convenient way to compare raw arithmetic throughput across processors and accelerators, though they do not capture all aspects of performance, such as latency, memory hierarchy, or parallel efficiency.