Decoupled access DRAM architecture

This paper discusses an approach to reducing memory latency in future systems. It focuses on systems where a single chip DRAM/processor will not be feasible even in 10 years, e.g. systems requiring a large memory and/or many CPU's. In such systems a solution needs to be found to DRAM latency and bandwidth as well as to inter-chip communication. Utilizing the projected advances in chip I/O bandwidth we propose to implement a decoupled access-execute processor where the access processor is placed in memory. A program is compiled to run as a computational process and several access processes with the latter executing in the DRAM processors. Instruction set extensions are discussed to support this paradigm. Using multi-level branch prediction the access processor stays ahead of the execute processor and keeps the latter supplied with data. The system reduces latency by moving address computation to memory and thus avoiding sending address to memory by the computational processor. This and the fetch-ahead capabilities of the access processor are combined with multiple DRAM "streaming" to improve performance. DRAM caching is assumed to be used to assist in this as well.