What are exaFLOPS?

The computing power of modern computers is measured in FLOPS, which stands for floating point op­er­a­tions per second. Very powerful systems can reach the level of exaFLOPS, which is a quin­til­lion (10¹⁸) op­er­a­tions per second. This enormous computing capacity is crucial for complex cal­cu­la­tions.

What are FLOPS and what are they used for?

FLOPS stands for floating point operations per second. Floating point arith­met­ic in computers is essential for sci­entif­ic cal­cu­la­tions, sim­u­la­tions, ar­ti­fi­cial in­tel­li­gence and other computing-heavy ap­plic­a­tions. Unlike simple op­er­a­tions with integers, floating point numbers make it possible to calculate and represent decimal numbers very precisely.

A computer’s per­form­ance in FLOPS is measured using spe­cial­ised bench­marks that test how many floating point op­er­a­tions can be performed per second. That often involves programs like LINPACK and HPCG, which perform complex math­em­at­ic­al cal­cu­la­tions and measure computing speed under real-world con­di­tions. Actual per­form­ance can vary based on hardware, software op­tim­isa­tion and the type of cal­cu­la­tion.

How many FLOPS are in an exaFLOPS?

One exaFLOP cor­res­ponds to exactly 10¹⁸ FLOPS or one quin­til­lion floating point op­er­a­tions per second. For com­par­is­on, one petaFLOPS (10¹⁵ FLOPS) is a thousand times smaller than an exaFLOPS, and one teraFLOPS (10¹² FLOPS) is a million times smaller.

What are some other FLOPS units and how are they converted to exaFLOPS?

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

These units make it easier to compare the computing power of different systems to each other.

How many FLOPS do modern computers and GPUs reach?

Modern high-per­form­ance computers now reach the levels of exaFLOPS. The first exaFLOPS su­per­com­puter was the Tianhe-3 in the Chinese National Su­per­com­puter Center in Guangzhou, which delivers up to 1.7 exaFLOPS. Other su­per­com­puters like the Frontier in the US now also operate at exaFLOPS levels.

High-end graphics cards like those used for gaming and machine learning also have high FLOPS values, even if they don’t come anywhere near exaFLOPS:

• NVIDIA RTX 4090: Around 82 teraFLOPS (0.000082 exaFLOPS) for FP32 cal­cu­la­tions; with over­clock­ing it can exceed 1,000 teraFLOPS
• NVIDIA H100 SXM: Around 989 teraFLOPS (0.000989 exaFLOPS) for FP32 Tensor Core cal­cu­la­tions
• NVIDIA A30: Around 10.3 teraFLOPS (0.000010 exaFLOPS) for FP32 cal­cu­la­tions
• AMD Radeon RX 7900 XTX: Around 61 teraFLOPS (0.000061 exaFLOPS) for FP32 cal­cu­la­tions

While GPUs are primarily designed for parallel cal­cu­la­tions in AI and graphics pro­cessing, su­per­com­puters are used for very complex sim­u­la­tions in science and industry.