High Performance and Energy Efficient Serial Prefetch Architecture

Energy efficient architecture research has flourished recently, in an attempt to address packaging and cooling concerns of current microprocessor designs, as well as battery life for mobile computers. Moreover, architects have become increasingly concerned with the complexity of their designs in the face of scalability, verification, and manufacturing concerns. In this paper, we propose and evaluate a high performance, energy and complexity efficient front-end prefetch architecture. This design, called Serial Prefetching, combines a high fetch bandwidth branch prediction and efficient instruction prefetching architecture with a low-energy instruction cache. Serial Prefetching explores the benefit of decoupling the tag component of the cache from the data component. Cache blocks are first verified by the tag component of the cache, and then the accesses are put into a queue to be consumed by the data component of the instruction cache. Energy is saved by only accessing the correct way of the data component specified by the tag lookup in a previous cycle. The tag component does not stall on a I-cache miss, only the data component. The accesses that miss in the tag component are speculatively brought in from lower levels of the memory hierarchy. This in effect performs a prefetch, while the access migrates through the queue to be consumed by the data component.