teraFLOPS

Teraflops, short for terafloating-point operations per second, is a unit of computing performance used to express the speed of processors, graphics units, and specialized accelerators. One teraflop equals 10^12 floating-point operations per second. In some contexts, teraflops are referenced as 2^40 flops (about 1.099 trillion) per second, though the decimal definition is standard for most engineering uses. The term is typically reported as peak theoretical performance, representing the maximum operations the hardware could perform under ideal conditions.

Performance figures are commonly given for single-precision (FP32) or double-precision (FP64) floating point operations. A device might offer a certain number of TFLOPS in FP32 and a different amount in FP64. Real-world performance often falls short of the peak due to memory bandwidth, software efficiency, and algorithm characteristics. Sustained FLOPS are generally lower than peak and vary with workload. TFLOPS figures scale with parallelism, so more cores and wider vector units enable higher FLOPS, particularly on GPUs and other throughput-oriented accelerators.

TFLOPS have become a standard metric in high-performance computing. The progression from gigaflops to teraflops accompanied advances in multi-core CPUs and GPUs, with the field advancing toward petaflops and exaflops in large systems. Milestones include early systems in the tens of gigaflops, the 2008 achievement of about 1 PFLOP with Roadrunner, and modern supercomputers measured in tens to hundreds of PFLOPS. While useful for comparison, FLOPS-based metrics are best understood alongside memory bandwidth, software efficiency, and application characteristics.