What Are FLOPS? Understanding AI Computing Performance.

What Are FLOPS?

FLOPS stands for floating-point operations per second, a measure of how many floating-point calculations a computing system can perform in one second.

Floating-point operations are mathematical calculations involving numerical values represented with a decimal-like range and varying precision.

FLOPS is commonly used when describing the performance of GPUs, supercomputers, and AI accelerators.

For example, a GPU manufacturer may advertise computing performance in teraFLOPS, representing trillions of floating-point operations per second under specified conditions.

Key Takeaways

• FLOPS measures the rate of floating-point calculations performed by computing hardware.
• Hardware performance may differ substantially depending on numerical precision and the operations being measured.
• FLOPS alone does not determine how quickly an AI model or robotics application will run.

How Are FLOPS Measured?

One FLOP is one floating-point operation.

FLOPS describes the number of those operations completed per second.

Common units include:

GFLOPS: Billions of operations per second.

TFLOPS: Trillions of operations per second.

PFLOPS: Quadrillions of operations per second.

For example, a processor capable of sustaining 20 trillion floating-point operations per second on a particular workload achieves 20 TFLOPS for that workload.

However, manufacturers often publish theoretical peak performance rather than performance measured in a complete application.

Why Does Numerical Precision Matter?

Different floating-point formats use different numbers of bits to represent values.

Common formats include FP32, FP16, and FP8.

A processor may execute low-precision calculations faster than higher-precision calculations, depending on its hardware.

For this reason, a GPU's advertised FP8 performance can be very different from its FP32 performance.

Comparisons must use the same precision and compatible assumptions.

Some advertised AI performance figures also include additional conditions, such as sparsity, which can substantially affect the numbers.

FLOPS vs. Real AI Performance

Consider two GPUs with similar theoretical computing performance.

One may have greater memory bandwidth, more suitable software support, or a better implementation of a particular neural network operation.

It may therefore run the intended AI model faster.

Actual model performance depends on memory movement, software efficiency, model architecture, and available computing resources.

In robotics, end-to-end latency can be more important than peak arithmetic throughput.

A perception model must produce useful results quickly enough to support the complete robot system.

FLOPS vs. FLOPs

The capitalization can matter.

FLOPS generally refers to a rate: operations per second.

FLOPs is sometimes used to describe the total number of floating-point operations required for a task.

For example, a model may require a certain number of operations for one inference, while a processor delivers a measured operation rate.

Frequently Asked Questions

Does twice the FLOPS mean twice the AI speed?

No. Performance depends on many factors beyond peak arithmetic capability.

Are TOPS and FLOPS the same?

No. TOPS measures operations per second more generally and may include integer operations. FLOPS specifically measures floating-point operations.

The Bottom Line

FLOPS provides a useful measure of computing capacity, but it must be interpreted alongside precision, memory performance, software, and workload characteristics. For Physical AI, real application performance matters more than the largest advertised number.