The fashionable benchmarking inferior Geekbench has launched a caller cross-platform tool to measure the show of devices nether AI-heavy workloads. Geekbench AI measures a device’s CPU, GPU, and NPU (neural processing unit) to find however good it tin grip instrumentality learning applications.
Geekbench developer Primate Labs has been moving connected the bundle utilizing the sanction Geekbench ML, which launched successful preview successful 2021 but shifted the sanction to AI for reasons that look obvious. To research however antithetic hardware responds to antithetic AI-related tasks, it evaluates show based connected some accuracy and speed, with enactment for antithetic frameworks, including ONNX, CoreML, TensorFlow Lite, and OpenVINO.
1/2
Screenshot: Tom Warren / The Verge
It delivers 3 scores, afloat precision, fractional precision, and quantized. Primate Labs says the scores besides person an accuracy measurement to measure however adjacent a workload’s outputs are to the truth, “or however accurately that exemplary tin bash what it’s expected to do.”
We’ll request much clip with devices moving on-device AI such arsenic Copilot Plus PCs and each of the caller phones to spot however show successful real-world tasks correlates with Geekbench AI’s numbers. Checking framerates oregon loading times is 1 happening — present we mightiness beryllium checking the accuracy of predictive text, oregon what generative AI-enabled representation exertion comes up with.
You tin download the instrumentality present to effort it yourself connected Windows, macOS, Linux, Android, and iOS.