Hybrid FPGA : architecture and interface

• Main Author: Yu, Chi Wai
• Other Authors: Leong, Philip
• Published: Imperial College London 2010
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.517695

Hybrid FPGAs (Field Programmable Gate Arrays) are composed of general-purpose logic resources with different granularities, together with domain-specific coarse-grained units. This thesis proposes a novel hybrid FPGA architecture with embedded coarse-grained Floating Point Units (FPUs) to improve the floating point capability of FPGAs. Based on the proposed hybrid FPGA architecture, we examine three aspects to optimise the speed and area for domain-specific applications. First, we examine the interface between large coarse-grained embedded blocks (EBs) and fine-grained elements in hybrid FPGAs. The interface includes parameters for varying: (1) aspect ratio of EBs, (2) position of the EBs in the FPGA, (3) I/O pins arrangement of EBs, (4) interconnect flexibility of EBs, and (5) location of additional embedded elements such as memory. Second, we examine the interconnect structure for hybrid FPGAs. We investigate how large and highdensity EBs affect the routing demand for hybrid FPGAs over a set of domain-specific applications. We then propose three routing optimisation methods to meet the additional routing demand introduced by large EBs: (1) identifying the best separation distance between EBs, (2) adding routing switches on EBs to increase routing flexibility, and (3) introducing wider channel width near the edge of EBs. We study and compare the trade-offs in delay, area and routability of these three optimisation methods. Finally, we employ common subgraph extraction to determine the number of floating point adders/subtractors, multipliers and wordblocks in the FPUs. The wordblocks include registers and can implement fixed point operations. We study the area, speed and utilisation trade-offs of the selected FPU subgraphs in a set of floating point benchmark circuits. We develop an optimised coarse-grained FPU, taking into account both architectural and system-level issues. Furthermore, we investigate the trade-offs between granularities and performance by composing small FPUs into a large FPU. The results of this thesis would help design a domain-specific hybrid FPGA to meet user requirements, by optimising for speed, area or a combination of speed and area.