Why PC Benchmarks Use the Geometric Mean Across 15+ Games

When evaluating computer hardware like graphics cards and processors, professional reviewers rely on a combination of a large sample size—typically 15 or more games—and a specific mathematical formula called the geometric mean. This article provides an overview of why this specific methodology is the industry standard. Testing at least 15 games ensures that hardware quirks, game engine biases, and API differences are smoothed out, while the geometric mean prevents games with naturally high frame rates from skewing the overall performance average. Together, these practices provide consumers with an accurate, unmanipulated single-number metric to compare the relative performance of different components.

The Problem with the Standard Average

To understand why the geometric mean is necessary, you first have to look at why the standard average (arithmetic mean) fails in hardware benchmarking. The arithmetic mean is calculated by adding up all the frame rates from every game tested and dividing that total by the number of games.

This approach is deeply flawed because different games run at vastly different frame rates. Consider a benchmark comparing two graphics cards across just two games: Counter-Strike 2 (an esports title) and Cyberpunk 2077 (a demanding AAA title).

  • GPU A gets 500 FPS in Counter-Strike and 50 FPS in Cyberpunk. The standard average is 275 FPS.
  • GPU B gets 400 FPS in Counter-Strike (20% slower) and 100 FPS in Cyberpunk (100% faster). The standard average is 250 FPS.

If a reviewer uses the standard average, GPU A wins with 275 FPS compared to GPU B's 250 FPS. However, this is misleading. GPU B literally doubled the performance in the demanding game while only losing by 20% in the esports title. Because Counter-Strike operates in the hundreds of frames per second, its raw numbers completely dominate the standard average, rendering the Cyberpunk results mathematically irrelevant.

How the Geometric Mean Solves the Problem

The geometric mean fixes the weighting issue of the standard average. Instead of adding the frame rates together, the geometric mean multiplies the results of all the games together and then calculates the nth root (where n is the number of games tested).

In practical terms, the geometric mean normalizes the data. It stops looking at raw frame rate totals and instead treats every game as an equal percentage weight.

Using the previous example, the geometric mean evaluates the ratios of performance. It recognizes a 10% frame rate increase in a game running at 60 FPS as having the exact same weight as a 10% increase in a game running at 600 FPS. By calculating the geometric mean differential across a test suite, reviewers can accurately state that one component is "X% faster" than another, without high-framerate esports titles distorting the final conclusion.

Why 15 or More Games is the Magic Number

Calculating the geometric mean is only half the battle; the data fed into the formula must also be comprehensive. Testing just three to five games is insufficient for a reliable hardware review because software varies wildly in how it utilizes computer components. Reviewers use 15 or more games to establish statistical significance and eliminate bias.

Here is why a large game sample size is required:

  • Developer and Sponsor Bias: Many games are developed in partnership with AMD, Nvidia, or Intel. A game optimized specifically for AMD architectures might perform poorly on equivalent Nvidia hardware. Testing 15+ games dilutes the impact of brand-sponsored titles.
  • Game Engine Diversity: Different engines (Unreal Engine 5, Frostbite, REDengine, id Tech) stress hardware differently. Some rely heavily on fast memory, while others care more about raw core counts or cache size.
  • Graphics APIs: Older games often use DirectX 11, while modern titles utilize DirectX 12 or Vulkan. Hardware architectures handle these APIs with varying degrees of efficiency.
  • Hardware Bottlenecks: A small sample size might accidentally hide a major flaw. For example, a graphics card with only 8GB of VRAM might perform brilliantly in 10 older games, but suffer massive performance penalties in 5 modern, texture-heavy games. A 15-game suite exposes these limitations.
  • Outlier Mitigation: Sometimes, a broken game patch or a flawed graphics driver causes a specific piece of hardware to underperform terribly in one specific title. In a 3-game test, that single anomaly ruins the whole review. In a 15-game test, the geometric mean absorbs the outlier, keeping the final performance summary accurate.

By combining a broad sample size of 15 or more diverse games with the normalizing math of the geometric mean, hardware reviewers protect consumers from misleading data. It is the only way to condense complex, wildly varying performance metrics into a single, trustworthy percentage differential.