POTDP: Research GPU Performance Optimization Method based on Thread Dynamic Programming

Abstract

GPU is widely used in high-performance computing such as big data and artificial intelligence because of its high concurrency and high throughput. With the development of VLSI technology, more and more processing units are integrated on chip. High power consumption increases the operating cost of equipment, reduces the battery life and reliability of integrated circuit chip, which seriously restricts the improvement of integrated circuit chip performance and restricts the expansion and application field of parallel systems. In view of the above problem, this paper proposes a data dependent GPU power management method– DDPM to reduce the power con-sumption of GPU system. The experimental results of DDPM show that compared with the shared aware data management method, DDPM improves the L1 cache hit rate by 2.8%, reduces DRAM data transmission capacity by 5%, and improves the average energy efficiency by 4.67% compared with MC-aware-ORI, MC-aware-LoSe and MC-aware-SiOb.