Running the Open Source Version of GFXBench on Android (2): Execution

This is the second installment in the series covering the GPU benchmarking software GFXBench.
In the previous article, we successfully built the GitHub-released GFXBench.
This time, we’ll actually run it and check the benchmark scores.

GFXBench consists of a variety of benchmarks that you select and run at startup.
The categories are broadly divided into High-Level Tests and Low-Level Tests.

In this article, we'll focus on the High-Level Tests and highlight a few of them.

There are many details described, but let’s focus solely on the “Required minimum API” section.
From it, you can see how the benchmarks have evolved alongside the OpenGL ES versions.
(It appears that each benchmark was added as GFXBench was upgraded.)

Therefore, the benchmarks you can run depend on the OpenGL ES version supported by your Android device. If a benchmark is not supported, it will be disabled on the benchmark selection screen.

※ Aztec Ruins seems to have taken a different direction, so it is omitted (it appears more like a multi-API benchmark rather than version-based).

Install the APK file built in the previous article onto your Android device.
Ensure that the Android device is in developer mode and connected to the build PC via adb.
When running adb commands, it’s better to use Windows Terminal instead of Git Bash to avoid (various) issues.

adb install gfxbench-5.1.5+corporate.apk

On your Android device, tap the GFXBench icon to launch it.
On the first launch, you will see a “Pushed data not found” message.

Select “OK” to copy the APK bundle data into the app’s data area.

Once launched, the benchmark’s main screen appears.

Tap the “Select Test” button to open the test selection screen.

For the benchmarks to run, I’ll choose the following two based on my personal interest:

• T-Rex
• Manhattan

Each also has an on-screen and an off-screen version.
The on-screen version renders the benchmark on the actual display, meaning the scores depend on the device’s screen resolution.
The off-screen version does not render to the screen (aside from minimal progress indicators) and runs at a fixed off-screen resolution, so the scores are independent of the device’s screen size.
If you want to compare GPU performance across different devices without being affected by screen resolution, choose the off-screen version.

When you first launch the selection screen, all benchmarks are checked by default, so uncheck everything except the benchmarks you want to run, then tap “Start”.
(It’s convenient to uncheck entire categories like High-Level Tests or Low-Level Tests in one go.)

Let’s run the off-screen versions.
(If you want to see the on-screen rendering, feel free to use the on-screen versions.)
I selected an Android device from my collection with a (relatively) recent Android version for the test.

Note that the devices we’re using for this article are in a spec range where off-screen benchmark scores peak in the mid-100 fps range.
(For high-end devices, the T-Rex and Manhattan benchmarks are now considered light…
In those cases, more demanding benchmarks would be more appropriate.)

Below are some sample measured results. The benchmark scores are the off-screen fps values.

Coincidentally :-) these GPUs differ by the number of cores (MC), and the ones with more cores produced better results (as expected).
To compare a bit further, let’s divide the benchmark scores by the number of cores.

From these values, you can see that the MC2 and MC3 devices follow the same trend (proportional to the assumed MC1 baseline).
By contrast, the MC1 device in our test shows a modest score.
It may be due to its lower specs, such as a lower clock frequency, among other reasons.

In this article, we introduced GFXBench’s features and actually ran the GFXBench build from the previous article to check the benchmark scores.
Next time, we’ll cover the procedure for running it on devices with older Android versions than those used here.

The screenshots and test results in this article use and reference the software and assets from Kishonti-Opensource/gfxbench, which is published under the BSD 3-Clause License.

• Original Copyright: (c) 2005–2025 Kishonti Ltd.
• License: BSD 3-Clause License
• Rights to Images, etc.: The copyright for the GFXBench benchmark screen images and UI quoted in this article belongs to the original author, Kishonti Ltd.

Disclaimer
The procedures, benchmark scores, and other test results presented in this article are provided on an AS IS basis, reflecting the state of a specific test environment, and their accuracy, safety, and reproducibility are not guaranteed.
Neither the author nor Mamezou Inc. shall be liable for any direct or indirect damages arising from the use of or reliance on this information. Please review the content carefully and use it at your own risk.