Exploring hybrid memory for GPU energy efficiency through software-hardware co-design
PACT '13 Proceedings of the 22nd international conference on Parallel architectures and compilation techniques
Hi-index | 0.00 |
Many core architectural computing devices such as Graphic Processing Unit (GPU) are becoming increasingly popular in scientific computing because of their performance advantages. The scaling trends of processor and memory technologies demand more innovations in memory and processor architectures, e.g., the need for new architectural techniques to leverage Non-Volatile Random Access Memories (NVRAM). However, there is a lack of versatile simulation tools for application programmers and system architects to explore the design space of GPU global memory organization and management. In order to fill this void, we build a new GPGPU simulator called vGPGPUSim that can construct GPU global memory with different compositions of memory devices and simulate its performance and power consumption. Our initial results have shown that vGPGPU-Sim can faithfully reveal the internal execution details of versatile GPU global memory.