High Performance VIM Vector Coprocessor Architecture And Implementation

PIM(Processor-In-Memory) merges processing and memory into a single chip to lower memory latency, increase memory bandwidth which brings a new way solving the well known "memory wall" problem. PIM plays an import role in modern architecture design. Vector model is naturally suitable for science computing, but its implementation cost keeps high so it is usually adopted by some main frames. VTM(Vector-In-Memory) architecture is based on PIM and vector technology. Vector processor based on VIM architecture is a cost effective solution in science computing and media processing etc domains.There are some differences in architecture between traditional vector processor and vector processor based on PIM technology. Studies have shown that though memory system is the main bottleneck in vector architecture, traditional in-order architecture can not keep the memory system always busy and it usually has high implementation cost.This thesis has focused on key technologies of vector processor based on VIM. Based on our study of traditional in-order, decoupled and out-of-order architectures, we proposed the VIM vector processor architecture, which adopts an improved decoupled architecture and distributed register file. VIM vector processor keeps implementation cost low and adds support to short vector instructions. Memory system has always been the import part of vector architecture and we have studied the new characters of memory system based on PIM technology. we proposed an improved DRAM access scheduling method, which introduces dynamic scheduling scheme in DRAM memory controller to improve DRAM memory bandwidth. And also in this thesis, we have implemented a vector microprocessor KD-VIM-1 based on the VIM architecture. This thesis also focused on implementation technology of vector microprocessor based on VIM architecture, including KD-VIM-1 ISA defination, multiported register file and vector execution pipeline etc.Our KD-VIM-1 vector microprocessor was implemented in VHDL hardware description language and finished synthesized into Altera Stratix series FPGA chips. KD-VIM-1 microprocessor's verification was based on the ProcSuperStar developing board.