UCLA

In the semiconductor industry today, ASICs are able to offer 10x-1000x higher energy and area efficiencies than non-dedicated chips, such as programmable DSP processers, field-programmable gate arrays (FPGAs), and microprocessors. Not surprisingly, SoCs today have become an integration of many ASIC blocks, each performing a few dedicated tasks. The growing size of modern SoC chips, accelerated by the increasing demands for functionalities, has exposed the major drawback of ASIC: design cost. These large SoCs are re-designed a few times a year to rectify hardware-bugs and to support new features. Because ASICs are not reconfigurable, even the smallest hardware change would require a re-design. Additionally, design cost is rising exponentially with every technology generation.

The rising design cost of ASICs has exposed a huge need today: efficiency and flexibility must co-exist. But among flexible hardware candidates, microprocessors and programmable DSP processors are far too slow to meet the throughput requirements of ASICs. FPGAs do come close in terms of performance, but are extremely inefficient due to its high energy and large area overhead. We must bridge the huge gap in efficiency for FPGA to become a viable contender to ASICs.

The primary culprit for FPGA inefficiency is interconnect, which accounts for over 75% of area and delay. For over 20 years, 2D-mesh network has been the back-bone of FPGA interconnects, but full connectivity in a 2D-mesh require O(N2) switches, requiring interconnects to grow much faster than Moore's Law. As a result, various heuristics are used to simplify switch-box arrays at the cost of resource utilization, but interconnect area of modern FPGA is still around 80%. This work builds FPGA using hierarchical interconnects based on Bene networks, requiring O(N∙log∙N) switches. Although Bene is commonly used in telecommunication, this work is its first silicon realization of a FPGA. To realize a highly efficient interconnect architecture, significant pruning of the network is required. Novel techniques such as fast-path U-turns and unbalanced branching are also implemented. A custom place-and-route software is developed to map benchmark designs on a variety of interconnect candidates. From mapping results, the architecture is updated based on network utilization until an optimized design is converged. The large area of FPGA chip requires aggressive power gating (PG), but interconnect signals often lack spatial locality, make it block-level PG difficult. A novel PG circuit technique is developed to power-gate individual interconnect switches with very small overhead in area and performance. Such technique requires fundamental circuit changes, even modifying the CMOS inverter.

With innovations in chip architecture, circuit design, and extensive software development, this work has demonstrated 5 user-mappable FPGAs (from 1K-16K LUTs) all with around 50% interconnect area: a 3-4x reduction from commercial FPGAs while preserving connectivity. An energy efficiency of 1.1 GOPS/mW is the highest among reported FPGAs, and is 22x more efficient than the most efficient commercial FPGA today, significantly bridging the efficiency gap between FPGA and ASIC.