GEMM Input Variations Dataset
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/theo-gregersen/input-dependent-power-sc24
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含了针对通用矩阵-矩阵乘法（GEMM）操作的输入数据变化，重点研究这些变化如何影响GPU的功耗。实验涵盖了多种数据类型（FP32、FP16、FP16-T、INT8），并探索了四种变化：数值分布、位相似性、布局和稀疏性。该数据集以2048×2048矩阵的规模进行测试，任务是在GEMM操作期间测量功耗。
This dataset covers variations in input data for general matrix-matrix multiplication (GEMM) operations, focusing on how these variations affect GPU power consumption. The experiments cover multiple data types (FP32, FP16, FP16-T, INT8) and explore four types of variations: numerical distribution, bit similarity, layout, and sparsity. The dataset is tested with matrices of size 2048×2048, and the task is to measure power consumption during GEMM operations.
