Geekbench is moving its artificial intelligence benchmark to prime time. Developer Primate Labs is pushing the newly renamed Geekbench AI (previously called 'Geekbench ML' while in preview) to version 1.0. The benchmark is available for Windows, Linux, and macOS, as well as for Android in the Google Play Store and for iPhones and iPads in Apple's App Store.
Like regular Geekbench, which provides two scores for the CPU, Geekbench AI will use multiple scores. In this case, the software will use three scores that range in precision: single-precision data, half-precision data, and quantized data. Primate Labs claims that this is "to better describe the multidimensionality of AI performance and the impact of different hardware designs."
But there's another component: correctness. Another measurement for each test sees how close the workload delivers what's expected. Primate Labs explained using an example of a model that detects the presence of a hotdog:
"Your hotdog object detection model might be able to run very, very quickly, but if it can only accurately detect a hotdog 0.2% of the time one is actually present, it’s not very good," it explained. "This accuracy measurement can also help developers see the benefits and drawbacks of smaller data types — which can increase performance and efficiency at the expense of (potentially!) lower accuracy. Comparing accuracy as part of performance through the use of our database can also help developers estimate relative efficiency."
