A Technology-Scalable Multithreaded Architecture
Clecio Donizete Lima, Kentaro Sano, Hiroaki Kobayashi, Tadao Nakamura, Michael J. Flynn · 2001
Advances in integrated circuit technology have offered an increasing transistor density with a continuous performance improvement. Soon, it will be possible to integrate a billion of transistors on a chip running at very high speeds. At this levei of integration, however, physical constraints related to wire delay will become dominant, requiring microprocessors to be more partitioned and use short wires for on-chip communication. On the other hand, effective parallel processing by taking advantage of the large number of transistors will be challenging. In this research, we propose the Shift Architecture, a multithreaded paradigm that maps statically scheduled threads onto multiple functional units. Communication is based on shift register files and restricted to contiguous functional units, requiring reduced wire lengths. Threads are dynamically interleaved on a cycle-by-cycle basis, to maintain high processor utilization. We describe the basic concepts of our approach. A preliminary evaluation shows that this architecture has the potential for achieving high instruction throughput for multithreaded benchmarks.