Simulation-based evaluation of the Imagine stream processor with scientific programs

This paper evaluates the performance and efficiency of Imagine stream processor for scientific programs. It classifies scientific programs into three classes based on their computation to memory access ratios. Typical programs are selected and measured on a simulated Imagine and two general-purpose processors. Comparative analysis is made with special attentions on the computational efficiency, power efficiency and performance bottleneck. The results show that programs in Class 1 attain high performance and high computational efficiency on Imagine. Their performance is only restricted by limited ILP and load imbalance across ALUs. Programs in Class 2 attain moderate performance but non-satisfactory computational efficiency on Imagine due to limited LRF and SRF reuse. Programs in Class 3 attain extremely low performance on Imagine because most of the execution time is spent on DRAM accesses. The estimated power efficiency on Imagine is excellent except for programs in Class 3.