Date of Award
Doctor of Philosophy (PhD)
Electrical Engineering and Computer Science
C.Y. Roger Chen
Behavioral models, Early prediction, Many-core computing, Signal Integrity, System-level design methodology
Electrical and Computer Engineering
Modern high performance computing applications such as personal computing, gaming, numerical simulations require application-specific integrated circuits (ASICs) that comprises of many cores. Performance for these applications depends mainly on latency of interconnects which transfer data between cores that implement applications by distributing tasks. Time-to-market is a critical consideration while designing ASICs for these applications. Therefore, to reduce design cycle time, predicting system performance accurately at an early stage of design is essential. With process technology in nanometer era, physical phenomena such as crosstalk, reflection on the propagating signal have a direct impact on performance. Incorporating these effects provides a better performance estimate at an early stage. This work presents a methodology for better performance prediction at an early stage of design, achieved by mapping system specification to a circuit-level netlist description.
At system-level, to simplify description and for efficient simulation, SystemVerilog descriptions are employed. For modeling system performance at this abstraction, queueing theory based bounded queue models are applied. At the circuit level, behavioral Input/Output Buffer Information Specification (IBIS) models can be used for analyzing effects of these physical phenomena on on-chip signal integrity and hence performance.
For behavioral circuit-level performance simulation with IBIS models, a netlist must be described consisting of interacting cores and a communication link. Two new netlists, IBIS-ISS and IBIS-AMI-ISS are introduced for this purpose. The cores are represented by a macromodel automatically generated by a developed tool from IBIS models. The generated IBIS models are employed in the new netlists. Early performance prediction methodology maps a system specification to an instance of these netlists to provide a better performance estimate at an early stage of design. The methodology is scalable in nanometer process technology and can be reused in different designs.
Deepaksubramanyan, Boray S., "EARLY PERFORMANCE PREDICTION METHODOLOGY FOR MANY-CORES ON CHIP BASED APPLICATIONS" (2014). Dissertations - ALL. 162.