BODCA: Heterogeneous CPU-GPU computing system with Bandwidth-Optimized DRAM cache design

摘要

Heterogeneous CPU-GPU computing is important due to multi-core CPU and GPU. In heterogeneous CPUGPU computing, cache coherence and unified memory spaces allow CPU and GPU cores to access the same memories without memory copy overheads. This paper focuses on the potential of the large-scale DRAM cache design in a heterogeneous computing simulation, considering the tradeoff of DRAM cache design. Also, this paper analyzes the improved latency and utilization of the system bandwidth of DRAM cache design. We propose a Bandwidth Optimized DRAM cache (BODCA), which is similar to Alloy Cache but has a dynamic bandwidth assignment technique. On average, BODCA shows 5.1% IPC improvement and 7.3% increment in the utilization of the memory system bandwidth compared to the Alloy Cache.