What are TFLOPS?

TeraFLOPS (TFLOPS) is a unit that indicates how many trillions of cal­cu­la­tions with floating point numbers a computer can carry out in one second. The value serves as a measure of the per­form­ance of pro­cessors, es­pe­cially GPUs and su­per­com­puters. TFLOPS are es­pe­cially relevant for ap­plic­a­tions that involve a lot of cal­cu­la­tion, like ar­ti­fi­cial in­tel­li­gence, sci­entif­ic sim­u­la­tions and machine learning.

What are FLOPS and what are they used for?

FLOPS stands for floating point operations per second and is a unit for computing power. A floating point operation is a math­em­at­ic­al cal­cu­la­tion that involves decimal points. They are es­pe­cially important for cal­cu­la­tion-heavy ap­plic­a­tions that require a high degree of precision.

FLOPS are mostly used for sci­entif­ic cal­cu­la­tions, sim­u­la­tions, ar­ti­fi­cial in­tel­li­gence, machine learning and graphics ap­plic­a­tions. They play a central role in various areas such as medical image pro­cessing and physical sim­u­la­tions. They are also important in finance, for example when it comes to the analysis of market data. In the gaming industry, FLOPS are used to determine the graphics per­form­ance of modern GPUs. With ever in­creas­ing FLOPS capacity, modern computers can deliver more and more realistic physical effects and high-res­ol­u­tion graphics.

FLOPS are typically measured using specially developed benchmark tests that determine the number of floating point op­er­a­tions per second. Fre­quently used bench­marks include LINPACK, which is mostly used for su­per­com­puters, and FP32/FP64, which rate the computing power of GPUs. During tests, complex math­em­at­ic­al cal­cu­la­tions are performed in order to determine how many op­er­a­tions per second a system can handle. Man­u­fac­tur­ers often give the­or­et­ic­al FLOPS values based on the ar­chi­tec­ture of a computer. However, real-world ap­plic­a­tions can vary based on workload and ef­fi­ciency.

How many FLOPS are in a teraFLOPS?

One teraFLOPS is equal to one trillion (1,000,000,000,000 or 10¹²) floating point op­er­a­tions per second. That means that a processor with 1 TFLOPS can execute a trillion math­em­at­ic­al op­er­a­tions with floating point numbers per second.

By way of com­par­is­on, a computer that only has 1 FLOPS would need 31,000 years to perform a trillion floating point op­er­a­tions. So computers working in TFLOPS are powerful systems capable of modern ap­plic­a­tions in real time.

What other FLOPS units exist and how do they convert into TFLOPS?

There are many FLOPS units, which differ in how many op­er­a­tions per second they refer to.

Su­per­com­puters’ per­form­ance is measured in petaFLOPS and even exaFLOPS, while high-end graphics cards are usually rated in teraFLOPS.

How many FLOPS do modern computers and GPUs reach?

GPUs and modern computers in the area of high-per­form­ance computing have reached im­press­ive FLOPS values. The NVIDIA H100, one of the most powerful GPUs for AI and data centres, achieves up to 989 teraFLOPS for FP32 Tensor Core cal­cu­la­tions. That makes it ideal for large neural networks and sim­u­la­tions.

The NVIDIA A30, a GPU that’s optimised for data centres, reaches 10 TFLOPS and is par­tic­u­larly suitable for AI training and in­fer­ences. By com­par­is­on, the gamer-oriented NVIDIA RTX 4090 can overclock to over 100 TFLOPS and enables very realistic graphics.

Su­per­com­puters are even more powerful: The Frontier su­per­com­puter has surpassed the 1 exaFLOPS mark and is used for highly complex sci­entif­ic sim­u­la­tions. Other powerful su­per­com­puters used in research, like the Japanese computer Fugaku, also operate in this range.