For some time now, processors have been offering cores specially focused on AI tasks. They are called NPUs and their capacity is measured in TOPS.
We see, in principle, that all processors on the market offer the same power in TOPS. Initially, one might tend to think that in this task they will all offer the same performance, more or less. But this is not the case and to make a proper comparison the Geekbench AI tool has been launched.
Measuring the AI performance of a processor is easy with Geekbench AI
First of all, you should know that this tool is not something new. Primate Labs has been offering Geekbench ML software for some time now to measure or evaluate the AI capabilities of processors. Actually, it is just a small name change, making it more user-friendly.
The term ML (Machine Learning), although more accurate, is not user-friendly. That is why it has been changed to AI (artificial intelligence). Pure marketing.
You should be aware that AI algorithms can perform very differently on different platforms. Because this software is cross-platform, users may see different scores depending on the system.
To provide accurate measurement and relevant information, the test results are divided into three groups: CPU, GPU and NPU. In fact, NPUs are a completely unnecessary addition. The GPU, especially the one integrated into processors, already offers good AI capabilities, due to its design focused on parallel calculations.
NPUs offer high inference speeds with low power consumption, in short. In addition, these cores are a prerequisite for using Microsoft’s Copilot+, which has recently been included in Windows 11.
Something quite interesting that Primate Labs explains is that AI benchmarks are more complex than GPU tests. It is not just raw processing power that should be measured. This tool shows three parameters, such as:
- Total precision
- Average precision
- Quantified scores
The idea behind this is to offer users a better perspective so they can make better choices based on the metrics. It would be something similar to what happens with processors, where performance is shown on a single core and on all cores.
This tool will undoubtedly evolve as the different models are updated and become more complex. Currently, AI algorithms are constantly evolving, and are still in fairly early stages of development.
If you’re interested, you can visit the official Geekbench AI ranking, but at the moment it only shows performance data, smartphones and tablets (mainly from Apple). We imagine that, now that the tool is a little better known, data on desktop processors will appear.