Gigaflops per Watt

Computer chips perform calculations and consume electricity while doing so. How much they compute is measured in flops: this is the number of calculation steps with decimal numbers per second. A gigaflop is a billion such steps per second. How much electricity a chip draws is measured in watts, just like with a light bulb. If you divide one by the other, you get gigaflops per watt. The figure answers a simple question: How much computing work do I get for a given amount of energy?

Why data centers stare at this number

A large data center for artificial intelligence can draw as much electricity as a small city. There, the power bill is often the largest ongoing cost item, larger than the technicians' salaries. Anyone who needs half the electricity for the same computing performance cuts a substantial part of their operating costs in half. That’s why efficiency isn’t a peripheral environmental issue for operators, but a hard business metric.

There’s also a physical limit. Every watt a chip consumes ultimately turns into heat. This heat has to be removed, otherwise the chip slows down for self-protection. A building has a fixed power connection and a cooling system with fixed capacity. Once this limit is reached, more performance can only be gained through better efficiency, not through more devices.

That’s why gigaflops per watt shows up in almost every product announcement from chip manufacturers. Manufacturers like Nvidia or AMD rarely advertise raw speed alone. They almost always also show how much performance per watt the new generation delivers compared to the old one. For customers, this is the decisive point of comparison.

What makes a chip more economical

The most important lever is manufacturing technology. The smaller the switching elements on the chip, the less energy a single calculation step costs. The figure given in nanometers, such as 5 or 3 nanometers, roughly describes this size. Smaller structures mean shorter paths for the current and therefore less loss.

The second lever is the precision of the numbers. A chip can store a number very precisely, with many decimal places, or more coarsely with few. Calculating with coarse numbers costs significantly less energy. For AI models, this coarser calculation is usually entirely sufficient, because the result barely gets worse. This is precisely why, when comparing figures, one must always check which precision a value applies to.

A common misconception is that an economical chip is automatically a fast chip. That’s not true. You can increase efficiency by letting the chip run at a lower clock speed, because consumption rises disproportionately with speed. A smartphone processor often has good efficiency but little absolute performance. Gigaflops per watt and absolute gigaflops are two different figures, and both matter.

Where the metric shows up in news and products

It’s most visible in the Green500, a ranking of the world’s most energy-efficient supercomputers. There, it doesn’t show who computes fastest, but who achieves the most per watt. The top values there now stand at several tens of gigaflops per watt. Fifteen years ago, it was a fraction of that.

The figure also plays a role in financial reporting. When corporations like Microsoft or Google pour billions into new data centers, the return depends directly on the efficiency of the chips installed. Analysts therefore calculate how many AI responses an operator can deliver per kilowatt-hour. Rising electricity prices make this calculation even more important.

In everyday life, one encounters the same principle under other names. If a laptop lasts longer despite the same battery size, an more efficient processor is almost always behind it. For smartphones, computing performance per watt is even more important than peak performance, because the battery is small and there are no fans.