Geekbench AI 1.0 released

The “Geekbench ML” gets a rename.

Primate Labs has released the first version of a benchmark called Geekbench AI. Although the name is new, the software is based on Geekbench ML, which has been available for a while and has been used for the newer AI release. Developers are aware that the term ML (Machine Learning) has been somewhat neglected, with the more marketing-friendly term AI (Artificial Intelligence) taking over, so Geekbench is following that trend.

AI algorithms can perform very differently on various platforms. Since Geekbench AI is cross-platform, users will see vastly different scores across tests. For this reason, the AI benchmark leaderboard will be divided into three groups: CPUs, GPUs, and NPUs (Neural Processing Units). The latter has been the focus for all processor makers, as they aim to ensure high inference speeds at low power for modern systems. Using a fast NPU is also a requirement for Microsoft’s Copilot+PC certifications.

The Geekbench AI benchmark will not check for certifications; instead, it will use its own algorithms to verify the speed of new processors and, in its usual fashion, display a score. This score can be compared between devices in each group, although nothing prevents someone from comparing a phone to a 450W graphics card.

Geekbench AI on GPU, Source: VideoCardz

Primate Labs explains that AI benchmarks are much more complex than GPU testing, as illustrated in an example. It’s not just raw computing power that needs to be measured. Over time, graphics vendors have added new frameworks and API support, which must also be reflected in benchmarking. The Geekbench GPU test focuses on various subtests, but they are hardly comparable to gaming benchmarks, which have become even more segmented with the availability of upscaling and frame generation. Sharing a single performance figure at the end of the test will never tell the whole story. The developer says this problem is even more complex for AI benchmarks.

In the end, Geekbench AI results will be divided into three sub-scores: full precision, half-precision, and quantized scores . This decision aims to provide better insights into how developers or hardware vendors interpret benchmarks, offering more dimensions, similar to how CPU testing revolves around single and multi-threaded scores. It seems very likely that AI benchmarking will evolve as algorithms become more complex.

Meanwhile, users can check the fastest AI hardware by browsing the official ranking . The official ranking may be a bit slow when adding new devices, so you can also check the browser with searchable functions that will provide more details and more recent tests. The Geekbench AI benchmark is available for free as a trial.

Source: Geekbench , Documentation (PDF)