Overhead

Measuring performance comes at a cost. In order to gather metrics, the system must interact with several subsystems — including CPU schedulers, GPU drivers, memory managers, and OS-level instrumentation interfaces — to collect, process, and store the data. These operations consume resources themselves, meaning that the act of profiling can introduce a small but measurable overhead on the device being tested. While GameBench is designed to minimise this impact, it is important to be aware that no performance measurement tool operates entirely without footprint, and results should be interpreted with this inherent trade-off in mind.

The measurements below were taken using a GameBench dummy application that places a constant, known load on both CPU and GPU, ensuring a reproducible baseline for comparing overhead across different profiling configurations.

Overhead on Android (11+)

Google Pixel 6 — Google Tensor (GS101), Mali-G78 MP20 GPU

CPU Overhead

	CPU busy	Overhead
Desktop v1.37 (baseline, not profiling)	24.4%
Desktop v1.37 (USB, all metrics)	28.1%	3.8%
Desktop v1.37 (USB, all metrics + device logcat)	27.5%	3.1%
Desktop v1.37 (USB, all metrics, screenshots disabled)	27.5%	3.1%
Desktop v1.37 (WIFI, all metrics)	28.8%	4.4%
Desktop v1.37 (WIFI, all metrics, screenshots disabled)	27.5%	3.1%

CPU busy is measured with “top”, normalized.

GPU Overhead

	GPU avg	GPU max	Median FPS
Desktop v1.37 (USB, all metrics)	63.40%	74.20%	41
Desktop v1.37 (USB, only GPU and FPS metrics enabled)	64.50%	75.00%	42

Overhead on Windows

Intel Core i7-9750H (2.6 GHz)

CPU Impact

• Typical overhead: 1–2% additional CPU usage
• Peak overhead: Up to 4% during intensive data collection periods

Memory Impact

• RAM usage: Approximately 60 MB additional memory consumption

GPU Impact

• Indirect impact: Screenshot capture uses Windows system APIs which may briefly utilize GPU resources