Geekbench AI benchmark introduced for testing performance in AI workloads

Neural processing units (NPUs), designed to accelerate artificial intelligence algorithms, were originally found on Apple’s mobile platforms and PC processors, but are now becoming commonplace in chips from Intel and AMD. This means that there is a need to evaluate their performance – the Geekbench AI test is designed to solve this problem.

Image source: geekbench.com

Primate Labs’ main Geekbench app is designed to test CPU and GPU performance, but in recent years the developer has also experimented with an alternative Geekbench ML. With the launch of Microsoft’s Copilot+ initiative and the start of the race for the title of the most powerful neural processors between Intel, AMD, Qualcomm and Apple, Primate Labs has released version 1.0 of a new test and renamed it Geekbench AI.

«Geekbench AI provides a proprietary suite of workload tests running on single-precision, half-precision, and quantized data, covering a variety of types that developers use for both precision and purpose in AI systems,” Primate Labs explained. Geekbench AI supports several AI frameworks: OpenVINO for Windows and Linux, ONNX for Windows, QNN for Qualcomm Snapdragon processors on PCs, Apple CoreML for macOS and iOS, as well as a number of Android frameworks offered by various chip manufacturers. On Windows PCs, the test supports neutral processors from Intel and Qualcomm – work with AMD will appear later.

Geekbench AI is available on Windows, macOS, Linux, iOS/iPadOS and Android. The basic version is free, but there is also a Pro license for $99 – it includes command line tools, the ability to run a test without uploading results to the Geekbench database, and some other features. The developers intend to update the benchmark as necessary, adding support for new hardware, frameworks and workloads.