Emulating large complex designs require multi-FPGA systems (MFS). However, inter-FPGA communication is confronted by the challenge of lack of interconnect capacity due to limited number of FPGA input/output (I/O) pins. Serializing parallel signals onto a single trace effectively addresses the limited I/O pin obstacle. Besides the multiplexing scheme and multiplexing ratio (number of inter-FPGA signals per trace), the choice of the MFS routing architecture also affect the critical path latency. The routing architecture of an MFS is the interconnection pattern of FPGAs, fixed wires and/or programmable interconnect chips. Performance of existing MFS routing architectures is also limited by off-chip interface selection. In this dissertation we proposed novel 2D and 3D latency-optimized time-multiplexed MFS routing architectures. We used rigorous experimental approach and real sequential benchmark circuits to evaluate and compare the proposed and existing MFS routing architectures. This research provides a new insight into the encouraging effects of using off-chip optical interface and three dimensional MFS routing architectures. The vertical stacking results in shorter off-chip links improving the overall system frequency with the additional advantage of smaller footprint area. The proposed 3D architectures employed serialized interconnect between intra-plane and inter-plane FPGAs to address the pin limitation problem. Additionally, all off-chip links are replaced by optical fibers that exhibited latency improvement and resulted in faster MFS. Results indicated that exploiting third dimension provided latency and area improvements as compared to 2D MFS. We also proposed latency-optimized planar 2D MFS architectures in which electrical interconnections are replaced by optical interface in same spatial distribution. Performance evaluation and comparison showed that the proposed architectures have reduced critical path delay and system frequency improvement as compared to conventional MFS. We also experimentally evaluated and compared the system performance of three inter-FPGA communication schemes i.e. Logic Multiplexing, SERDES and MGT in conjunction with two routing architectures i.e. Completely Connected Graph (CCG) and TORUS. Experimental results showed that SERDES attained maximum frequency than the other two schemes. However, for very high multiplexing ratios, the performance of SERDES & MGT became comparable.
Experimental Evaluation and Comparison of Time-Multiplexed Multi-FPGA Routing Architectures
2017-10-05
Electronic Theses and Dissertations
Theses
Electronic Resource
English
DDC: | 629 |
Ultrafast Time-Division Multiplexed Multi-Access Networks
British Library Conference Proceedings | 1996
|Comparison of multi-UAV cooperation architectures
IEEE | 2017
|