| Increasing power density with technology scaling has caused stagnation in operating frequency of modern day microprocessors. This has led designers to prefer multi-core architectures over complex monolithic processors to keep up with the demand for rising computing throughput. Although processing units are getting smaller and simpler, the dramatic rise of their count on a single die has made the fabric that connects these processing units increasingly complex. These interconnect fabrics have become a bottleneck in improving overall system efficiency. As a result, the design paradigm for multi-core chips is gradually shifting from a core-centric architecture towards an interconnect-centric architecture, where system efficiency is limited by the fabric rather than the processing ability of any individual core. This dissertation introduces three novel and synergistic circuit techniques to improve scalability of switch fabrics to make on-die integration of hundreds to thousands of cores feasible. 1) A matrix topology is proposed for designing a fully connected switch fabric that reuses output buses for programming, and stores shuffle configurations at cross points. This significantly reduces routing congestion, lowers area/power, and improves performance. Silicon measurements demonstrate 47% energy savings in a 64-lane SIMD processor fabricated in 65nrn CMOS over a conventional implementation. 2) A novel approach to handle high radix arbitration along with data routing is proposed. It optimally uses existing cross-bar interconnect resources without requiring any additional overhead. Bandwidth exceeding 2Tb/s is recorded in a test prototype fabricated in 65nm. 3) Building on the later, a new circuit topology to manage and update priority adaptively within the switch fabric without incurring additional delay or area is then proposed. Several assist circuit techniques, such as a thyristor based sense amplifier and self regenerating bi-directional repeaters are proposed for high speed energy efficient signaling to and from the switch fabric to improve overall routing efficiency. Using these techniques a 64 x 64 switch fabric with 128b data bus fabricated in 45nm achieves a throughput of 4.5Tb/s at single cycle latency while operating at 559MHz. |
