Study of GPU scalability

General-purpose Graphic Processing Units (GPGPUs) have been widely used to accelerate heavy compute-intensive applications. In a market the number of GPU cores on one chip are increased to maximize performance by exploiting thread-level parallelism (TLP). But memory system could be a bottleneck, and assigning more cores does not help faster execution. In this paper, we observe GPU core scalability issue, and analyze major causes of the performance degradation under many core configuration in perspective of memory latency.