Geekbench has an AI benchmark now

The popular benchmarking utility Geekbench has launched a new cross-platform tool to evaluate the performance of devices under AI-heavy workloads. Geekbench AI measures a device’s CPU, GPU, and NPU (neural processing unit) to determine how well it can handle machine learning applications.

Geekbench developer Primate Labs has been working on the software using the name Geekbench ML, which launched in preview in 2021 but shifted the name to AI for reasons that seem obvious. To explore how different hardware responds to different AI-related tasks, it evaluates performance based on both accuracy and speed, with support for different frameworks, including ONNX, CoreML, TensorFlow Lite, and OpenVINO.

It delivers three scores, full precision, half precision, and quantized. Primate Labs says the scores also have an accuracy measurement to evaluate how close a workload’s outputs are to the truth, “or how accurately that model can do what it’s supposed to do.”

We’ll need more time with devices running on-device AI such as Copilot Plus PCs and all of the new phones to see how performance in real-world tasks correlates with Geekbench AI’s numbers. Checking framerates or loading times is one thing — now we might be checking the accuracy of predictive text, or what generative AI-enabled image editor comes up with.

You can download the tool now to try it yourself on Windows, macOS, Linux, Android, and iOS.

https://www.theverge.com/2024/8/15/24221382/geekbench-ai-benchmark-software