High Performance Cache Architectures to Support Dynamic Superscalar Microprocessors

Simple cache structures are not sufficient to provide the memory bandwidth needed by a dynamic superscalar computer, so more sophisticated memory hierarchies such as non-blocking and pipelined caches are required. To provide direction for the designers of modern high performance microprocessors, we investigate the performance tradeoffs of the combinations of cache size, blocking and non-blocking caches, and pipeline depth of caches within the memory subsystem of a dynamic superscalar processor for integer applications. The results show that the dynamic superscalar processor can hide about two-thirds of the additional latency of two and three pipelined caches, and that a non-blocking cache is always beneficial. A pipelined cache will only outperform a non-pipelined cache if the miss penalty and miss rates are large.