Professors Ma Chunyuan and Zhang Tao from Northwestern Polytechnical University, along with masters Chen Jing and Yao Ding, recently published a paper titled “Evaluation Methods for Computing Power of Embedded Intelligent Computers” in the Journal of Computer Science. The abstract of the paper is as follows. Scan the QR code or click “Read the original text” to view the full text.
The evaluation method for computing power is one of the research hotspots in the field of embedded intelligent computers. Embedded intelligent computers have various optimized solutions for neuromorphic computing and machine learning accelerators, making their evaluation more challenging than that of general-purpose computers. Benchmark testing is currently the most commonly used evaluation method; however, on resource-constrained embedded devices, the reusability of benchmark test sets and evaluation metrics is limited, making it difficult to adapt to the diverse configurations of embedded intelligent systems. Additionally, the computational intensity of neural network models in the test set has a certain degree of randomness, which does not fully exploit the computational potential of the devices under test. The evaluation metrics are not standardized, making it difficult to conduct comparative analyses of the computing power of different embedded intelligent computers. This paper proposes an evaluation method for the computing power of embedded intelligent computers based on a neural evolutionary algorithm. Firstly, based on the Roofline theory model, we integrate the advantages of computational potential mining, resource adaptation, and standardized evaluation metrics, proposing a framework for evaluating computing power that is adaptable to various embedded intelligent computers, and analyzing its rationality. Secondly, we propose a neural network model generation algorithm for evaluating computing power, utilizing a neural evolutionary algorithm to ensure that the computational intensity of the generated model approaches the upper limit of the computational intensity of the embedded intelligent computer, fully mining the computational potential of the devices under test, thus making the evaluation results more objective. Then, using a fixed environment host computer as a control, we run the generated neural network model crosswise on both the device under test and the host computer, taking the number of floating-point operations per second during two inference tasks as the computational factor, and providing a universal formula for computing power evaluation that allows for comparative analysis of the computing power of different embedded intelligent computers. Finally, we evaluated the Huawei Atlas 200 under the Mindspore-cpu, Tensorflow-cpu, and Mindspore-ascend310 frameworks. Compared to the five commonly used neural network models in benchmark tests, the evaluation results using the neural network models generated by this paper are more reasonable, confirming that the intelligent computing power of the two DaVinci cores is 42.37 times that of the eight Cortex-A55 cores.
Article QR code