Accelerate Dense Matrix Multiplication on Heterogeneous-GPUs
Jianan Sun, Mingxue Liao, Yongyue Chao, Pin Lv · 2023
Matrix multiplication is crucial in scientific computing, but it demands substantial resources. We propose a framework for effectively utilizing heterogeneous GPUs to large matrix multiplication. By splitting matrices into small blocks and using Douglas’s variant of Strassen’s algorithm, we enable concurrent tasks on heterogeneous systems. Our framework improves speed by 89.5% on homogeneous GPU servers and by 108% in multi-server heterogeneous GPU setups.