Design of a parallel vector access unit for SDRAM memory systems

Parallel Vector Access is a technique that exploits the regularity of vector or stream accesses to perform them efficiently in parallel on a multi-bank memory system. The performance of applications that have vector accesses may be improved using a memory controller that performs scatter/gather operations so that only the vector or stream elements that are accessed by the application are transmitted across the system bus. These scatter/gather operations can be speeded up by broadcasting vector operations to all banks of memory in parallel, each of which implements an algorithm to determine which elements of the requested vector they contain. This thesis presents the mathematical foundations behind one such algorithm for controller are investigated. The the performance of such a memory controller on vector kernels is studied by gate level simulation and the results analyzed. Because of the parallel approach, the PVA is able to load elements up to 32.8 times faster than a conventional memory system and 3.3 times faster than a pipelined vector unit, without hurting normal cache line fill performance.