Nowadays, almost every phone is being touted as AI enabled in a bid to get more customers, but it seems like that wouldn’t be a tagline for many phones again as Antutu, has released a dedicated benchmarking tool targetting AI-powered phones.
Antutu announced this new development via a blog post since the normal benchmarking tool can’t test for AI capabilities.
“Although AI is strictly an algorithm, it is inseparable from software and hardware support. The superposition of the two brings a performance gap. In order to let everyone have an intuitive judgment on the AI performance of their mobile phones, Ann Bunny officially released the “An Bunny AI Review” public beta, providing a quantifiable standard for everyone to judge the difference in AI performance of different platforms. “ Antutu wrote
Antutu also made it known that since there is no unified standard for AI and with different AI enabled chip out there with different approach and SDK, the Abb Bunny AI tool will test chips in two categories; “namely image classification and object recognition.”
“Among them, the image classification is based on the Inception v3 neural network, the test data is 200 pictures; and the object recognition is based on the mobilenet ssd neural network, the test data is a 600-frame video.” Antutu added.
An Bunny AI Grading criteria:
To make grading fair, here is how the AI tool works;
1. The score is related to the speed and accuracy. The faster the speed, the higher the accuracy and the higher the final score.
2. if the speed is faster, but the accuracy is poor, Ann Bunny AI evaluation has targeted penalties to avoid cheating behaviors that reduce the accuracy to increase the speed and ultimately affect the total score, and vice versa.
Special case description:
1. because the bunny AI evaluation is an evaluation of AI computing power, so the AI processor is the same, the score difference performance is not obvious (for example: Qualcomm 845 and 710 DSP models are the same, so the score gap is not large)
2. currently Samsung has not yet The release of its own AI SDK, so Samsung’s own chips are tested by the CPU, resulting in low results, this situation will be improved after Samsung released the SDK.
3. HiSili HiAI engine object recognition Currently using TFLite to perform calculations on the CPU, resulting in low scores, this situation will also be improved after HiSilicon upgrade HiAI.
4. NVIDIA’s mobile chip can pass the AI algorithm to the GPU through the TensorRT engine through floating point algorithm.
5. Android version will also have an impact on test scores, in theory, the same chip in Android 9.0 will be higher than Android 8.0, because Google has been optimizing the support of AI at the system level.