TOPS

TOPS is a unit of measurement for computing speed. The abbreviation stands for “Tera Operations Per Second,” meaning trillions of computing steps per second. A chip with 40 TOPS accordingly performs 40 trillion simple computing operations in one second. What’s meant here are very basic steps, primarily multiplications and additions of numbers. An AI program needs exactly these kinds of steps in enormous quantities when it recognizes an image or generates a sentence. Processor manufacturers therefore use TOPS as a shorthand formula for how “AI-capable” their hardware is.

The number chipmakers advertise with

TOPS has become the most important marketing figure for processors in recent years. For graphics cards, it used to be the clock speed in gigahertz; today, for AI chips, it’s the TOPS value. Microsoft has even made the unit a requirement: a laptop may only call itself a “Copilot+ PC” if its AI chip delivers at least 40 TOPS. This means a single number determines which product label a device is allowed to carry.

Economically, this reflects a larger shift. Until now, AI applications almost always ran in data centers, far away from the user. If chips on-site deliver enough TOPS, the computation can happen directly on the device. This saves server costs, works without internet, and keeps private data at home. For providers like Qualcomm, Apple, AMD, or Intel, this is a billion-dollar market.

Still, one should read the figure with caution. TOPS is a theoretical maximum under ideal conditions, not a measured result from real-world use. Whether a program actually achieves that performance depends on software, memory, and cooling. A car with a top speed of 300 km/h doesn’t drive any faster than any other car in city traffic.

What’s behind the number

The figure results from a simple multiplication. You take the number of compute units on the chip, multiply it by the operations per clock cycle, and by the clock frequency. The result is the theoretical peak performance. Because modern AI chips run thousands of small compute units in parallel, double-digit TOPS figures quickly emerge.

One often-omitted detail is crucial: the precision of the numbers. A chip can compute with coarse numbers, such as whole values between 0 and 255, or with finer decimal numbers. Coarse numbers are faster, so the TOPS value increases when precision is lowered. A chip with 50 TOPS at coarse precision may only manage 25 or less at finer precision. Reputable specifications therefore always state the format alongside the figure, usually as “INT8” or “INT4”.

Related to this is the unit FLOPS, which counts computing steps involving decimal numbers. Large training systems are almost always specified in FLOPS, while small device chips are specified in TOPS. A direct comparison of the two figures doesn’t really make sense, even though it’s often drawn in articles.

TOPS in phones, laptops, and cars

You most commonly encounter this unit in product descriptions for notebooks and smartphones. There, it refers to the NPU, the part of the chip built exclusively for AI tasks. A high-end smartphone currently sits at around 40 to 50 TOPS, a laptop chip at 40 to 80. This performance is sufficient for noise suppression in video calls, offline translations, or cutting out people from images.

The auto industry also calculates in TOPS. Systems for automated driving need to analyze camera images in real time, and manufacturers cite figures in the range of several hundred TOPS for this. In stock market announcements, the number then serves as proof that a supplier is technically ahead.

As a buyer, you shouldn’t overrate the number. It’s useful for a rough comparison within the same chip generation and the same number format. But it says nothing about whether your favorite software actually uses the NPU at all. Many programs still prefer to run on the processor or graphics chip – in which case even the most impressive TOPS figure goes unused.