Why do people talk about FLOPs when discussing AI models?

FLOPs give a rough, hardware-neutral way to describe how much computation training or running a model takes. That helps compare models, estimate cost and discuss how compute needs are growing.

Do I need to understand FLOPs to use AI in my business?

Not in detail. It helps to know that more compute usually means more cost and time, so you can ask vendors sensible questions about price per task and whether a smaller model would do.

AI Glossary / AI foundations / Beginner

Floating-point operations (FLOPs)

Q: What is the difference between FLOPs and FLOPS?

FLOPs with a lowercase s is a count of floating-point operations, a measure of how much work a task needs. FLOPS with a capital S is operations per second, a measure of how fast hardware can work.

Definition

Floating-point operations (FLOPs) are individual arithmetic steps, such as adding or multiplying two decimal numbers, that a computer performs. In AI, the total number of FLOPs is used as a measure of how much computation a task needs, for example training or running a model. The lowercase form FLOPs counts operations. FLOPS with a capital S usually means operations per second, a speed rating for hardware.

What it means

A floating-point number is a number that can have a decimal point, like 3.14 or 0.0007. Every time a computer adds, subtracts, multiplies or divides two such numbers, that is one floating-point operation. AI models are built almost entirely out of this kind of arithmetic, so counting FLOPs gives a rough, hardware-neutral way to describe how much work a job involves.

For a business reader, FLOPs are a unit of effort, a bit like kilowatt-hours for electricity or labor-hours for a project. A bigger FLOP count generally means more time, more chips and more cost. The count says nothing about whether the result is useful. Two models can use the same compute and perform very differently.

How it works

Inside a neural network, most of the work is multiplying and adding large tables of numbers. A single step often multiplies one input by one stored weight and adds the result to a running total. That pair is commonly counted as two operations. Multiply the number of such steps by the number of inputs processed and you get a FLOP estimate.

Two figures matter in practice. The first is total FLOPs for a job, such as one full training run of a model. The second is the hardware's peak speed, quoted in FLOPS, meaning operations per second. Dividing the first by the second gives a best-case time. Real jobs run slower than that, because chips spend time waiting for data to move in memory, and utilization of the rated peak is often well below one hundred percent.

The precision of the numbers also changes the picture. Many AI chips use lower-precision formats that take less memory and run faster, so a FLOP is not always the same size of task across different hardware. Published FLOP counts are usually estimates, and the exact way of counting (for instance whether a multiply and an add count as one or two) varies by source.

Three practical examples

Finance team: A finance team wants to run a model that reads thousands of invoices each night. Their vendor quotes the work in FLOPs per invoice. The team multiplies that by their volume to compare a rented cloud setup against buying a server.
Retailer: A retailer compares two demand-forecasting models. The larger one needs about ten times the FLOPs per forecast but improves accuracy only slightly. Because the retailer pays for compute by the hour, FLOPs help them see that the extra accuracy is not worth the extra cost.
Support desk: A support desk uses an AI assistant to draft replies. Longer customer messages mean more tokens to process, which means more FLOPs per reply. The team notices that trimming pasted email history from each request lowers both response time and usage charges.

Think of it this way

Think of FLOPs as the number of individual hand calculations a huge room of accountants would have to do to finish a job. More calculations mean more people, more hours and more cost. The analogy stops working because accountants check meaning as they go, while FLOPs only count arithmetic steps and say nothing about whether the answer is good.

Worked example

Suppose a small fictional model needs 4 billion FLOPs to process one customer message. A company handles 50,000 messages a day. Total daily work is 50,000 x 4,000,000,000 = 200,000,000,000,000 FLOPs, which is 200 trillion FLOPs.

Now suppose the server they rent is rated at 10 trillion FLOPS (10 trillion operations per second) at peak. Best-case time is 200 trillion / 10 trillion = 20 seconds of pure computation for the whole day's messages.

Real systems rarely hit peak. If the server reaches 25 percent of its rated speed, the effective rate is 10 trillion x 0.25 = 2.5 trillion FLOPS. The time becomes 200 trillion / 2.5 trillion = 80 seconds. The lesson: the FLOP count tells you the size of the job, and utilization tells you how far reality sits from the best case. Here the compute itself is tiny, so other costs such as software, storage and staff time will likely dominate.

Illustrative case study

Fictional case: Larkspur Logistics, a made-up regional freight broker, planned to add an AI tool that summarizes shipping documents. A vendor proposed a very large model. Before signing, an analyst asked for the FLOPs per document and the rated FLOPS of the hardware, then multiplied by expected monthly volume. The numbers showed the large model would need a dedicated and expensive server running most of the day.

The analyst then asked for the same figures for a smaller model. It needed roughly one fifth of the compute per document. The team ran a two-week test on 300 real documents and had staff score the summaries. The smaller model scored almost the same on accuracy. The problem the team had hit was assuming that bigger always meant better. The fix was to pick the smaller model, set a monthly compute budget, and review accuracy scores each quarter so they could upgrade only if the gap grew.

Common misconceptions and mistakes

More FLOPs always means a better model. Compute is one input among several. Data quality, design and training method also drive results, and a smaller model trained well can beat a larger one trained poorly.

FLOPs and FLOPS are the same thing. Lowercase FLOPs counts operations in a task, while FLOPS with a capital S is operations per second. Mixing them up leads to wrong time and cost estimates.

A hardware FLOPS rating tells you real-world speed. The rating is a peak under ideal conditions. Memory limits, software and data movement usually keep actual speed well below it.

FLOPs are only a concern for companies that build models. Anyone paying for AI by usage, tokens or compute time is paying indirectly for FLOPs, so the idea helps with budgeting and vendor comparison.

Business use

Most businesses will never count FLOPs directly. Start by asking vendors for compute cost per task in plain terms, such as cost per thousand documents or per hundred conversations, and ask how that connects to the model's size. If you run your own workloads, estimate FLOPs per job, multiply by volume and compare against the rated and realistic speed of your hardware.

Measure cost per task, response time, and accuracy side by side, not compute alone. The main risks are paying for far more model than the task needs, trusting peak speed ratings, and ignoring energy and cloud charges that grow with volume. FLOP figures from vendors are often estimates and counting methods differ, so treat them as a way to compare options roughly rather than as exact billing predictions.

Check your understanding

In the worked example, how many FLOPs does the company need per day?

Answer: 50,000 messages x 4 billion FLOPs each = 200 trillion FLOPs per day.

Why did the worked example take 80 seconds instead of 20?

Answer: The server only reached 25 percent of its rated peak speed, so the effective rate was 2.5 trillion FLOPS instead of 10 trillion, and 200 trillion / 2.5 trillion = 80 seconds.

What is the difference between FLOPs and FLOPS?

Answer: FLOPs counts operations in a job, while FLOPS measures operations per second of hardware.

Why did Larkspur Logistics choose the smaller model?

Answer: It needed about one fifth of the compute per document and scored almost the same on accuracy in a real test, so the larger model's extra cost was not justified.

Related terms

Compute
GPU
Training
Inference
Parameters
Token
Scaling laws
Latency

Learn next

Try estimating cost per task for one AI tool you already use, using the vendor's usage data and the simple multiply-and-divide steps in the worked example. Then read about compute, inference and parameters to see how model size, usage and cost connect.

Sources

Google, Machine Learning Glossary: https://developers.google.com/machine-learning/glossary Goodfellow, Bengio and Courville, Deep Learning: https://www.deeplearningbook.org/ Stanford HAI, AI Index: https://hai.stanford.edu/ai-index The sources were consulted for the standard definition of floating-point operations and for context on how compute is discussed in AI; the examples, analogy, worked example and case study are original educational illustrations. Checked 5 Nov 2026.

Previous: Fligner-Killeen test Next: Flow-based clustering