Memory subsystem design for explicit multithreading architectures

Explicit multithreading (XMT) is a parallel programming approach for exploiting on-chip parallelism. An important enabler for XMT is sufficient memory bandwidth to support parallelism. For targeted deep-submicron VLSI processes, chip designers will be faced with the widely acknowledged issues of rising inter-connect RC delays and shortening clock periods. Comprehensive memory design for an XMT architecture has never before been rigorously studied. This thesis relies on an examination the implications of the XMT programming model on memory subsystem design to motivate a potential framework for on-chip memory interconnection. Many system-level issues are considered, and analytical electrical interconnect modeling is used to demonstrate the physical viability of new structures in future processes. It is estimated that a chip, built in a 2008 process with 1024 hardware execution contexts, may be capable of a sustained on-chip memory transaction throughput of 1430 GB/s.