面向人工智能领域的芯片测评结果数据集

本数据使用S60芯片驱动提供的topsexec工具将多个主流人工智能模型及其输入数据分别进行封装，为每个模型生成一个.bin文件。该数据是一种衍生性的、结构化的科学数据，其使用的主流人工智能模型来自于自研生产的标准ONNX格式的模型数据和公开论文，模型输入数据则是由Pytorch的标准库函数生成的随机张量和模型附带的静态文件。其中，原始的主流人工智能模型的数据时间范围在2012年至2018年，空间范围是全随机数据，空间精度是多尺度张量，计算方式是图像分类、掩码语言建模、点击率预测、强化学习、目标监测等。在控制数据质量方面，（1）对比测试封装前的ONNX模型和封装后的.bin数据保证数据一致性；（2）所有输入数据按模型要求进行归一化、填充或编码保证输入数据标准化；（3）在S60芯片上运行.bin文件，与CPU/GPU上原始模型的测试结果对比验证，保证数据的正确性；（4）.bin文件包含模型结构、权重及输入数据的完整信息，支持重复加载与验证。该数据为 S60 平台提供即用型模型测试包以加速边缘 AI 部署流程，作为硬件推理速度、内存占用及能效比的统一基准测试套件，提供开箱即用的 AI 模型演示资源，适用于教学或技术展示，促进 S60 平台模型工具链的完善，降低开发者移植与调试成本。该数据集的数据量1.62GB。

This dataset packages multiple mainstream artificial intelligence models and their input data separately using the topsexec tool provided by the S60 chip driver, generating a dedicated .bin file for each model. This is a derivative, structured scientific dataset. The mainstream AI models used are derived from self-developed standard ONNX-format model data and public academic papers, while the model input data includes random tensors generated by standard library functions of PyTorch and static files attached to the models.

The original mainstream AI model data spans from 2012 to 2018, with full random spatial scope, multi-scale tensor spatial precision, and covers computation modes including image classification, masked language modeling, click-through rate prediction, reinforcement learning, object detection, etc.

For data quality control: (1) Conduct a comparative test between the pre-packaging ONNX model and the packaged .bin data to ensure data consistency; (2) Normalize, pad or encode all input data according to model requirements to guarantee input standardization; (3) Run the .bin files on the S60 chip and compare the test results with those of the original models on CPU/GPU to verify data correctness; (4) The .bin files contain complete information including model structure, weights and input data, supporting repeated loading and verification.

This dataset provides a ready-to-use model test package for the S60 platform to accelerate the edge AI deployment workflow, serving as a unified benchmark test suite for hardware inference speed, memory occupancy and energy efficiency ratio. It also offers out-of-the-box AI model demonstration resources suitable for teaching or technical exhibitions, helping to improve the S60 platform model toolchain and reduce the development costs of model transplantation and debugging. The total data size of this dataset is 1.62 GB.