AI Benchmark

Ignatov Andrey

About this app

Neural Image Generation, Face Recognition, Image Classification, Question Answering...

Is your smartphone capable of running the latest Deep Neural Networks to perform these and many other AI-based tasks? Does it have a dedicated AI Chip? Is it fast enough? Run AI Benchmark to professionally evaluate its AI Performance!

Current phone ranking: http://ai-benchmark.com/ranking

AI Benchmark measures the speed, accuracy, power consumption and memory requirements for several key AI, Computer Vision and NLP models. Among the tested solutions are Image Classification and Face Recognition methods, AI models performing neural image and text generation, neural networks used for Image / Video Super-Resolution and Photo Enhancement, as well as AI solutions used in autonomous driving systems and smartphones for real-time Depth Estimation and Semantic Image Segmentation. The visualization of the algorithms’ outputs allows to assess their results graphically and to get to know the current state-of-the-art in various AI fields.

In total, AI Benchmark consists of 83 tests and 30 sections listed below:

Section 1. Classification, MobileNet-V3
Section 2. Classification, Inception-V3
Section 3. Face Recognition, Swin Transformer
Section 4. Classification, EfficientNet-B4
Section 5. Classification, MobileViT-V2
Sections 6/7. Parallel Model Execution, 8 x Inception-V3
Section 8. Object Tracking, YOLO-V8
Section 9. Optical Character Recognition, ViT Transformer
Section 10. Semantic Segmentation, DeepLabV3+
Section 11. Parallel Segmentation, 2 x DeepLabV3+
Section 12. Semantic Segmentation, Segment Anything
Section 13. Photo Deblurring, IMDN
Section 14. Image Super-Resolution, ESRGAN
Section 15. Image Super-Resolution, SRGAN
Section 16. Image Denoising, U-Net
Section 17. Depth Estimation, MV3-Depth
Section 18. Depth Estimation, MiDaS 3.1
Section 19/20. Image Enhancement, DPED
Section 21. Learned Camera ISP, MicroISP
Section 22. Bokeh Effect Rendering, PyNET-V2 Mobile
Section 23. FullHD Video Super-Resolution, XLSR
Section 24/25. 4K Video Super-Resolution, VideoSR
Section 26. Question Answering, MobileBERT
Section 27. Neural Text Generation, Llama2
Section 28. Neural Text Generation, GPT2
Section 29. Neural Image Generation, Stable Diffusion V1.5
Section 30. Memory Limits, ResNet

Besides that, one can load and test their own TensorFlow Lite deep learning models in the PRO Mode.

A detailed description of the tests can be found here: http://ai-benchmark.com/tests.html

Note: Hardware acceleration is supported on all mobile SoCs with dedicated NPUs and AI accelerators, including Qualcomm Snapdragon, MediaTek Dimensity / Helio, Google Tensor, HiSilicon Kirin, Samsung Exynos, and UNISOC Tiger chipsets. Starting from AI Benchmark v4, one can also enable GPU-based AI acceleration on older devices in the settings ("Accelerate" -> "Enable GPU Acceleration" / "Arm NN", OpenGL ES-3.0+ is required).

Ratings and reviews

This is a very interesting app that showcases the different tasks the NPU has to deal with. Very intuitive. One gripe though, my 2 year old S21 Ultra scored 734.9k while my brand-new S23 Ultra scored a measly 132.6k. Both were running the latest of their respective softwares (Android 13 One UI 5.0/5.1). Anybody rocking an S23 with similar results? I should point out that the S23 did score higher in the accuracy and CPU. NN-Speed was WAY higher on the S21 than that of the S23 (48-337 vs 16-16).

Stops working on Samsung... Trying to compare with OnePlus but it bugged out... Useless.

Seems like it'd work if I had a better phone, but half way through the test it said "Unfortunately, AI Benchmark has stopped". I don't know if it's my phone, or the app it self. I'll be willing to give a higher rating if it gets fixed. My phone is a Samsung J7 Sky pro. edit: I just realized it's supposed to crash, sorry! I'll give it a 5 star now lol. it's a good app.

What's new

1. New tasks and models: Vision Transformer (ViT) architectures, Large Language Models (LLMs), Stable Diffusion network, etc.
2. Added tests checking the performance of quantized INT16 inference.
3. LiteRT (TFLite) runtime updated to version 2.17.
4. Updated Qualcomm QNN, MediaTek Neuron, TFLite NNAPI, GPU and Hexagon NN delegates.
5. Added Arm NN delegate for AI inference acceleration on Mali GPUs.
6. The total number of tests increased to 83.