TOPS stands for trillions of operations per second. It is a theoretical measurement of how many AI-related calculations a processor or AI accelerator can perform in one second.

On an AI PC, TOPS is commonly used to describe the performance of the Neural Processing Unit, or NPU. For example, an NPU rated at 40 TOPS can theoretically perform up to 40 trillion operations per second under the manufacturer’s specified test conditions.

However, TOPS does not tell you how fast every AI application will run. Real performance also depends on calculation precision, memory bandwidth, software support, drivers, model size and whether the application can use the NPU at all.

What Does TOPS Stand For?

TOPS means tera operations per second. One tera operation represents one trillion operations.

The keyword here is theoretical.

TOPS represents the peak processing potential of the hardware. It does not guarantee that an application will continuously reach that level.

The calculation method also matters. A TOPS rating may be measured using INT8, INT4 or another numerical format. Some specifications may include sparsity, while others do not. For this reason, two processors with similar TOPS ratings may deliver different results in the same application.

Intel describes TOPS as a measure of theoretical peak AI performance and notes that its platform figures are commonly calculated with INT8 operations. Its published examples also separate NPU, GPU and CPU contributions rather than treating them as identical processors. See Intel’s AI PC explanation.

What Is an NPU?

An NPU, or Neural Processing Unit, is a specialized processor designed to accelerate machine-learning and AI inference workloads.

Modern AI PCs may contain three different processing engines:

A CPU can run AI software, but it is designed for many different kinds of work.

A GPU can process many calculations in parallel and is often better suited to demanding generative AI workloads. However, it may consume more power than an NPU.

An NPU is designed to run supported AI workloads efficiently without occupying the CPU or GPU. This is especially helpful for continuous background features such as camera effects, voice enhancement and real-time recognition.

NPU TOPS vs Total TOPS

One of the most confusing parts of an AI PC specification is the difference between NPU TOPS and total system TOPS.

They are not the same measurement.

NPU TOPS

NPU TOPS measures only the theoretical AI performance of the dedicated neural processor.

This is the number that matters when:

• checking whether a PC meets an NPU requirement;
• comparing supported Windows AI features;
• evaluating an application designed specifically for the NPU;
• checking Copilot+ PC eligibility.

GPU AI TOPS

GPU AI TOPS describes AI processing potential from the graphics processor.

The GPU may be more useful than the NPU for:

• local image generation;
• larger language models;
• GPU-supported creative software;
• workloads that do not have an NPU execution path.

CPU AI performance

The CPU may also contribute to an AI workflow. It can prepare data, run unsupported operations and handle the application surrounding the AI model.

However, CPU AI performance should not be treated as equivalent to dedicated NPU performance.

Total system TOPS

Total TOPS usually combines the theoretical AI capabilities of the CPU, GPU and NPU.

A simplified product specification may look like this:

This does not mean that every application can use all 39 TOPS at the same time.

Software must be designed to distribute work across the available processors. If an application only supports the GPU, the NPU may remain unused. If it requires an NPU, GPU TOPS cannot automatically replace the missing NPU performance.

Therefore:

Do not compare total system TOPS with another product’s NPU-only TOPS.

Why Do Copilot+ PCs Need 40 TOPS?

Microsoft defines Copilot+ PCs as a category of Windows 11 devices with a dedicated NPU capable of more than 40 TOPS, alongside other hardware and software requirements.

The NPU is used for compatible on-device features such as:

• Live Captions with translation;
• Windows Studio Effects;
• AI-enhanced Windows search;
• Recall, where available;
• supported local Windows AI models.

Microsoft specifically refers to a 40+ TOPS NPU, not 40 total system TOPS. Feature availability may also depend on the processor platform, Windows version, region and software updates. See the official Microsoft Copilot+ PC requirements.

A computer with a 16 TOPS NPU and 39 total platform TOPS should therefore not be described as having a 39 TOPS NPU.

It may still support compatible local AI acceleration, but the two specifications answer different questions.

Is Higher TOPS Always Better?

Not necessarily.

A higher TOPS rating indicates greater theoretical processing potential under the stated conditions. It does not guarantee that one computer will be faster in every AI task.

Actual performance depends on several additional factors.

1. Software support

An application must be able to send its AI workload to the NPU.

If the software only supports the CPU or GPU, a 50 TOPS NPU may sit idle while the application is running.

Before purchasing an AI PC, check whether the software you use supports:

• the processor manufacturer;
• the specific NPU architecture;
• Windows ML or DirectML;
• ONNX Runtime;
• the vendor’s AI development tools.

2. Numerical precision

AI hardware can process data in formats such as INT4, INT8, FP16 and FP32.

Lower-precision calculations generally require less memory and can produce a higher operations-per-second figure. However, not every model can use every precision level without affecting compatibility or output quality.

A 50 TOPS INT4 figure should not automatically be compared with a 50 TOPS INT8 figure.

3. Memory capacity and bandwidth

The processor needs fast access to model weights and input data.

An AI accelerator may have a high theoretical TOPS rating but still be limited by:

• insufficient system memory;
• slow memory bandwidth;
• data-transfer delays;
• a model that is too large for the available memory.

This is especially important in compact PCs that use shared system memory for the CPU, integrated GPU and NPU.

4. Model compatibility

Different AI models contain different operations. Some can run efficiently on a particular NPU, while others may include unsupported operations that fall back to the CPU or GPU.

The model may therefore run, but not entirely on the NPU.

5. Sustained power and cooling

Peak TOPS does not necessarily describe sustained performance.

Power limits and cooling can influence how long a processor maintains its highest operating frequency. This is one reason why the complete computer design matters, not just the processor specification.

6. Latency and workload size

Small, real-time tasks may benefit from low latency and high efficiency. Larger models may depend more on memory capacity and GPU performance.

The best processor for background noise reduction may not be the best processor for generating an image or running a local language model.

TOPS vs TFLOPS vs Tokens per Second

TOPS is not the only AI performance metric.

TOPS is useful for understanding the potential of an NPU, but tokens per second and latency may be more meaningful when evaluating a local language model.

Likewise, TFLOPS can help describe GPU compute potential, but it does not directly predict performance in every AI application.

Does ChatGPT Use Your Computer’s NPU?

The cloud version of ChatGPT primarily runs its AI models on remote servers.

Your computer still handles the browser, interface, networking and other local tasks, but a higher local NPU TOPS rating does not automatically make a cloud-generated response arrive faster.

An NPU becomes more relevant when:

• the AI model runs locally;
• Windows assigns a supported feature to the NPU;
• an installed application includes an NPU execution provider;
• a developer specifically targets the NPU.

If most of your AI usage happens through cloud services, TOPS should not be the main reason to replace an otherwise suitable computer.

Is 16 TOPS Enough for an AI PC?

A 16 TOPS NPU can still be useful. It may accelerate compatible background and local AI workloads more efficiently than a CPU.

However, 16 NPU TOPS is below Microsoft’s 40+ TOPS requirement for the Copilot+ PC category.

Whether it is enough depends on your intended use:

A lower-TOPS system may still be a better computer for a particular user if it has more memory, stronger graphics, better storage expansion or more suitable ports.

A Real Example: NPU TOPS and Total AI Performance

The KAMRUI G2 Pro Mini PC provides a useful example of how AI specifications should be read.

Its product specifications list:

• an AMD Ryzen 9 270 processor;
• an integrated AMD Ryzen AI engine;
• up to 16 TOPS of NPU performance;
• up to 39 TOPS of combined CPU, GPU and NPU performance;
• Radeon 780M integrated graphics;
• 32GB of LPDDR5 memory.

The important distinction is that 16 TOPS refers to the NPU, while 39 TOPS refers to combined platform performance.

The 39 TOPS figure should not be compared directly with Microsoft’s 40+ NPU TOPS Copilot+ requirement.

At the same time, the complete system may still be relevant to users who value a balance of CPU performance, integrated graphics, local AI acceleration, multiple NVMe drives and external-GPU expansion.

This is why buyers should evaluate the entire configuration rather than choosing a computer from one AI number.

How Many TOPS Do You Need?

There is no universal TOPS requirement for every AI user.

Start with these questions:

1. Which AI applications will you use?
2. Do they run locally or in the cloud?
3. Do they support the NPU in the computer?
4. Do you specifically need Copilot+ features?
5. Is the workload limited by compute performance or memory?
6. Would it run better on the GPU?
7. How much RAM and storage does the complete workflow require?

If you cannot identify an application that uses the NPU, paying more solely for a higher TOPS figure may provide little immediate benefit.

If you need a supported Copilot+ feature, the NPU requirement becomes much more important.

How to Compare AI PCs

Before buying an AI PC, check the following:

• Is the advertised number NPU TOPS or total TOPS?
• What numerical precision was used?
• Does the figure include sparsity?
• Does your software support the NPU?
• Does the system meet the complete Copilot+ requirements?
• How much RAM is installed?
• Is the memory upgradeable?
• What GPU is included?
• Can the storage be expanded?
• Are the required ports available?
• Does the manufacturer provide current drivers?
• Are there independent application-specific results?

TOPS should be one item on this checklist—not the entire checklist.

Final Answer: What Does TOPS Mean in AI?

TOPS measures the theoretical number of trillions of operations an AI processor can perform per second.

It is useful for describing the potential of an NPU, GPU or complete AI platform, but it is not a universal measure of real-world AI speed.

When comparing computers:

• compare NPU TOPS with NPU TOPS;
• check the calculation precision;
• do not confuse total TOPS with dedicated NPU TOPS;
• confirm that your applications support the hardware;
• consider memory, GPU performance, cooling and software compatibility.

The best AI PC is not necessarily the one with the largest TOPS number. It is the one whose hardware and software match the AI workloads you actually plan to run.

Frequently Asked Questions

Is 40 TOPS good for AI?

A 40+ TOPS NPU meets the performance threshold Microsoft uses for the Copilot+ PC category. It can be suitable for supported on-device Windows AI features, but real performance still depends on software compatibility, memory and the complete processor platform.

Is 16 TOPS enough for AI?

It can be enough for supported light and background AI workloads. However, a 16 TOPS NPU does not meet Microsoft’s 40+ TOPS Copilot+ requirement.

Is NPU TOPS the same as total TOPS?

No. NPU TOPS measures the dedicated neural processor. Total TOPS may combine the theoretical AI performance of the CPU, GPU and NPU.

Does a higher TOPS number make games faster?

Not automatically. Most game performance depends more directly on the CPU, GPU, memory and game engine. An NPU may support specific AI features, but its TOPS rating is not a conventional gaming-performance metric.

Does ChatGPT use my NPU?

The cloud version of ChatGPT primarily runs on remote servers. Your NPU is more relevant to AI models and features that execute locally and explicitly support the NPU.

What is the difference between TOPS and TFLOPS?

TOPS usually describes low-precision AI operations, while TFLOPS measures floating-point operations. Both are theoretical compute metrics, but they describe different calculation types and cannot always be compared directly.

Does 39 total TOPS qualify a PC as Copilot+?

Not if the dedicated NPU is below Microsoft’s 40+ TOPS NPU requirement. Total platform TOPS and NPU TOPS are different specifications.

Can a GPU have more AI TOPS than an NPU?

Yes. A GPU may offer greater theoretical AI compute, especially for demanding workloads. An NPU is usually designed to provide better efficiency for supported, continuous on-device AI tasks.