面向通用HPC的高性能DSP设计权衡

Design Tradeoffs of High Performance DSPs for General-Purpose HPC

GPU由于其计算能力高达数TFLOPS,被高性能计算领域用于加速并行运算.但GPU较低的峰值性能利用率和功耗效率,已经成为了系统性能进一步提升的瓶颈.为了解决这个问题,作者开始研究将高性能DSP用于通用高性能计算领域.为了高效支撑通用高性能计算,文中提出了高性能DSP的结构框架,并通过映射GotoBLAS库到该结构上,建立了GEMM在该结构上的性能模型.作者研究了影响GEMM效率的主要因素,包括性能、存储层次、核的大小以及核的数量.文中总结了一些有指导意义的结论用于构建面向通用高性能计算的高效DSP.实验结果表明,通过尽可能少的硬件代价,可以在TFLOPS DSP上获得接近峰值的性能.

The traditional HPC area employs GPUs which can afford TFLOPS level computing ability to accelerate the parallel computing.The low peak performance utilization and the low power efficiency of GPUs have become the bottlenecks for the system performance improvement.We start introducing high performance DSPs into general-purpose HPC area to address this issue.To support general-purpose HPC effectively,this paper constructs a performance model for the GEMM on high performance DSPs by mapping GotoBLAS onto the proposed architecture.We investigate factors that influence the performance and efficiency of GEMM,including performance,memory hierarchy,core size and number of cores.Some suggestive conclusions are summarized to help designing DSPs that are efficient for the general-purpose HPC.Evaluation results show that it can achieve a near-peak performance on the TFLOPS DSP with as few hardware cost as possible.