@techreport{oai:ipsj.ixsq.nii.ac.jp:02003140,
 author = {Lingqi,Zhang and Jiajun,Huang and Sheng,Di and Satoshi,Matsuoka and Mohamed,Wahib and Lingqi Zhang and Jiajun Huang and Sheng Di and Satoshi Matsuoka and Mohamed Wahib},
 issue = {3},
 month = {Jul},
 note = {Tensor Cores are specialized units integrated in modern GPUs, designed to accelerate dense matrix operations with remarkable efficiency. They have proven particularly effective in compute-bound workloads, such as those found in deep learning training, where general matrix-matrix multiplication (GEMM) is prevalent. Motivated by this success, recent efforts have explored extending Tensor Core usage to non-GEMM computational patterns. However, despite their potential, effectively utilizing Tensor Cores in broader contexts requires a thorough understanding of their performance characteristics across diverse workloads. This work investigates the applicability of Tensor Cores to non-GEMM workloads, seeking to answer a fundamental question: Can Tensor Cores accelerate non-GEMM kernels?, Tensor Cores are specialized units integrated in modern GPUs, designed to accelerate dense matrix operations with remarkable efficiency. They have proven particularly effective in compute-bound workloads, such as those found in deep learning training, where general matrix-matrix multiplication (GEMM) is prevalent. Motivated by this success, recent efforts have explored extending Tensor Core usage to non-GEMM computational patterns. However, despite their potential, effectively utilizing Tensor Cores in broader contexts requires a thorough understanding of their performance characteristics across diverse workloads. This work investigates the applicability of Tensor Cores to non-GEMM workloads, seeking to answer a fundamental question: Can Tensor Cores accelerate non-GEMM kernels?},
 title = {Can Tensor Cores Accelerate Non-GEMMWorkloads? An Analytical Study},
 year = {2025}
}