Evaluation of Existing Architectures in IRAM Systems

Computer memory systems are increasingly a bottleneck limiting application performance. IRAM architectures, which integrate a CPU with DRAM main memory on a single chip, promise to remove this limitation by providing tremendous main memory bandwidth and significant reductions in memory latency. To determine whether existing microarchitectures can tap the potential performance advantages of IRAM systems, we examined both execution time analyses of existing microprocessors and system simulation of hypothetical processors. Our results indicate that, for current benchmarks, existing architectures, whether simple, superscalar or out-of-order, are unable to exploit IRAM's increased memory bandwidth and decreased memory latency to achieve significant performance benefits. 1 Introduction One proposed solution to the growing gap between microprocessor performance and main memory latency is to integrate a processor and DRAM on the same die, an organization we refer to as Intelligent RAM (IRAM) [..