There is now a benchmark out there for testing AI workloads. Brought out by Primate Labs, Geekbench AI is an additional piece of kit now alongside the more well-known Geekbench software. But now it’s got a simple solution to check out your device’s AI readiness. Based on its other offering of ML (machine learning) it has adapted that to something that’s more recognizable and now booming with the AI naming scheme.
With such a wide range of hardware capable of running these tasks, Geekbench AI is rather flexible on how you can run it. With support for macOS, Windows, Linux, Android, and iOS you can check pretty much any device. Along with any capable component on board you can choose the back end to target either CPU, GPU, or NPU with a separation of categories for easier comparisons and tests.
It also should help review and test some of those Copilot+ PCs that require an NPU on board with a certain amount of power behind it. Along with any hardware claiming the same thing without it, something like the MSI Prestige 14 AI Evo we reviewed and wondered how well a CPU-only laptop can actually do in any sort of AI workloads. So it’s certainly a good addition to our test suite.
Geekbench AI leaderboards and features
One of the highlights of having a good simple benchmark is being able to compare your hardware to others. Such as in Cinebench or 3DMark, you can easily see how yours stands against others, or if you’re looking to pick one up, be able to see which one ranks better for the job. In that case, Geekbench just gives you a score for your benchmark performance to get it on the leaderboard.
There you can see all of the recent results and additions across the board, where you can compare the likes of a 6900XT vs a Samsung Galaxy S22 Ultra as seen in our screenshot below. After a while, these should go through to the ML benchmarks table where the hardware is ranked against each other.
There is also a PDF of the workloads Geekbench AI uses that you can delve into. These are a lot more specific to what AI is used for and not a general use of hardware. As we see computer vision and processing workloads are included. The addition of image classification, face detection, and object detection then is subjected to text classification. All of these are real-world scenarios for the tech and are a good way to test them.
