Why GPU utilization shows N/A on some Android devices

GPU utilization is the share of time the graphics processor is busy over a given period. If the GPU spends half of each second working and half idle, its utilization is 50%. Per-app GPU utilization goes a step further: it attributes that busy time to a specific app, separating your game or benchmark from everything else drawing on the screen.

Both figures are useful, which is why people expect to see them. But producing an honest number requires the system to measure GPU activity and make that measurement available to apps. On Android, the second part is where things usually break down.

The operating system schedules work onto CPU cores itself, so it naturally keeps track of how much CPU time each process uses. An app can read its own figures and turn them into a usage percentage. That is how MobiBench can show the app's own CPU usage in the system monitor and during the CPU stress test.

The GPU works differently. Apps send rendering commands through graphics APIs such as OpenGL ES or Vulkan, and the GPU driver, written by the chip vendor, decides how and when that work runs on the hardware. The detailed knowledge of how busy the GPU is lives inside that driver and the hardware's own counters.

There are a few reasons why an ordinary app usually can't read per-app GPU utilization:

No standard public API

Android does not provide a standard, public API that reports GPU busy percentage to apps in the way the system tracks CPU time. There is no single call an app can make that works the same way across devices.

Vendor-specific interfaces

Some device and chip makers expose GPU load or performance counters through their own kernel or vendor-specific interfaces. These differ between GPU families, drivers and software versions, and they typically report activity for the whole GPU rather than for one app.

Security restrictions

Android's security model limits which system files and interfaces ordinary apps can read. Interfaces that expose hardware details are often among those restricted. Developer profiling tools can show GPU counters on supported devices, but they generally rely on special permissions or a connection to a computer, which a normal app from the Play Store doesn't have.

It is tempting to estimate GPU utilization from other numbers, such as frame rate or frame time. The trouble is that those numbers don't determine utilization.

• A low frame rate can mean the GPU is saturated, but it can also mean the CPU is the bottleneck while the GPU waits.
• When rendering is synchronised with the display, a GPU that finishes each frame early sits idle until the next refresh. The same FPS can come from a lightly loaded GPU or a heavily loaded one.
• Different GPUs do different amounts of work per frame, so no formula converts FPS into a percentage that holds across devices.

An estimated percentage would look precise while possibly being wrong, and comparing such figures between devices would be meaningless. Showing N/A is the honest answer when no legitimate measurement exists.

A missing number is better than a made-up one.

The MobiBench GPU test focuses on what can be measured directly: it renders and animates a detailed 3D brain-hologram model for about 60 seconds and reports FPS, minimum, maximum and average FPS, frame time, temperature, peak temperature and a GPU score based on the measured rendering performance. On the results screen, the render-time chart is labelled N/A utilization, so it is always clear which value is missing and why.

If you want a sense of how hard the GPU is working without a utilization figure, look at how the results change over the run. Steady frame times with a temperature that levels off suggest the device is sustaining the workload comfortably. Frame times that creep up while the temperature keeps rising suggest it is reaching its thermal limits and lowering its clocks.

Frame rate and frame time are often the more useful numbers anyway, because they describe what you actually see. Our guide to FPS and frame time explains how to read them. You'll also find a short answer about N/A in the MobiBench FAQ.