This article was 1st published on our sister Site, Digital World Native.
This free, open-source smartphone app, developed by Zurich-based team of computer vision researchers, determines whether your phone is fast enough & has the capability to run Artificial Intelligence (AI) tasks that use the latest deep neural networks. Phone makers, AI app developers, AI nerds, & just plain users can run the benchmark on any phone, where several key AI tasks are performed & compiled into a handy score.
The AI-Benchmark benchmarking app was developed by this group of Switzerland-based researchers & can be downloaded for free here on the Google Play Store. It measures speed & memory requirements for a number of key computer vision AI algorithms, which are also indicative of performance in other AI realms.
There a total of 9 tasks measured, which are performed by 9 separate neural networks running on the phone. A bunch of architectures are considered in order to assess performance & limits of the approaches used by each manufacturer when solving AI problems.
The following are the 9 tasks tested:
Task 1: Object Recognition / Classification
Task 2: Object Recognition / Classification (4x larger size than Task 1)
Task 3: Face Recognition
Task 4: Image Deblurring
Task 5: Image Super-Resolution
Task 6: Image Super-Resolution (with 2 neural networks)
Task 7: Semantic Image Segmentation
Task 8: Photo Enhancement
Task 9: Memory Limits (processing large size HD images)
Currently, the most important factors driving benchmark scores for phones are mobile processors & RAM size/speed. The HiSilicon Kirin 970 NPU rules the roost with AI-Benchmark scores of upwards of 6,519 on a Huawei P20 Pro, followed closely by this same NPU on other Huawei phones. Other phone manufacturers near the top are OnePlus, Google, HTC, Asus, Samsung, LG & Xiaomi.
Other processors that rank well on the benchmark ranking list are the Mediatek Helio P60, the Snapdragon 845, the Intel Core i5-7Y57 & the Exynos 9810 Octa. Phone models that provide native acceleration support for float or quantized AI models using NNAPI are by far the outstanding performers. The rankings are available on the project Website.
