Abstract
GPGPU-Sim nowadays has become an important vehicle for academic architecture research. In the aspect of machine learning, it has now been widely used in various applications, such as auto-drive, mobile device, and medication, etc. As these machine learning applications are power-consuming, which has become a critical issue in the machine learning area. This paper proposes the implementation of fixed-point instructions and enabled flow on GPGPU-Sim to replace floating-point instructions in machine learning applications which is with scalable precision. Preliminary experimental results with our revised GPGPU-Sim models show that this design saves GPU energy consumptions by 11% on average when using 16-bit fixed-point as the data type.