SCIMA: A Novel Architecture for High Performance Computing

Technological trends have brought the growing disparity between processor and memory speeds. This memory wall problem is becoming very serious especially in high performance computing. In this paper, we propose a new architecture SCIMA for solving this problem. In SCIMA, addressable memory is integrated into the processor chip besides ordinary cache. Since the on-chip memory is software controllable, it has more ability to make good use of data locality than data cache which is controlled by hardware. The purpose of on-chip memory is to reduce the off-chip memory traffic by exploiting data reusability as much as possible within a chip. We have evaluated SCIMA by using QCD simulation, a practical application in quantum field theory. The performance evaluation reveals that SCIMA successfully reduces off-chip memory traffic and achieves higher performance than cache-only processor.