GPU Testing Methodology

A result belongs to one workload

GPU Lab runs versioned browser workloads, not a universal graphics-card score. Geometry, textures, ray marching, compute, and video exercise different paths. Particle simulation includes CPU work. A result describes the selected path under the recorded settings; it does not predict every game or application.

Warm-up and timing methods

Warm-up frames are excluded from the summary. Frame intervals describe browser delivery and can include scheduling or CPU delays. GPU timer queries, when available and valid, describe GPU work separately. Compute completion wall time includes submission and synchronization. Always read the timing-method label before comparing numbers.

Repeat a matching configuration

Standard benchmark mode fixes the workload settings. Compare the workload version, actual API, resolution, preset, duration, and timing method. Keep browser version, power mode, display conditions, and background tasks consistent. Repeat at least three runs and consider their spread, rather than selecting the fastest result.

Reports preserve context

Reports record requested and actual APIs, fallback reasons, workload version, settings, samples, and the stop reason. A stopped or hidden-page run is incomplete evidence. Custom shaders and local video are not standard comparison workloads. Saved results do not certify temperature, utilization, physical memory integrity, or hardware health.

• Open GPU Benchmark
• About the publisher