Abstract
Low-power miniature systems for ubiquitous computing such as wireless sensor networks have been developing rapidly in the past years. The growing demand on collecting and analyzing information from surrounding environment drives researchers and engineers to develop Internet-of-Things (IoT). This trend requires future integrated circuits for IoT devices to be ultra-low-power (ULP), flexible, and low-cost. Existing circuit solutions of IoT devices are either too costly such as sub-threshold ASICs, or too power-consuming such as sub-threshold microprocessors. ULP FPGAs operating in near/sub-threshold region, flexible and much lower-power than sub-threshold microprocessors, become a promising hardware solution for IoT applications. In this dissertation, circuit/architecture and tool flow of a custom ULP FPGA are explored and developed.
1) Energy Efficient FPGA Interconnect
The global interconnect is the major power consumer of the core fabric of FPGAs. Studies have shown that over 65% of power is dissipated in the interconnection fabric. The same conclusion applies to delay and area. The strict requirements on both speed and energy of IoT applications make energy reduction and energy-efficiency improvement of FPGA routing fabrics a driving challenge. In this dissertation, an energy-efficient low-swing interconnect is modeled, optimized, and evaluated in near/sub-threshold region. When implementing Microelectronics Center of North Carolina (MCNC) benchmarks, the proposed interconnect leads to 68.4% delay reduction and 47.5% energy reduction compared to prior works.
2) Per-Path Voltage Scaling and Power-Gating
Per-path voltage scaling is a technique to reduce FPGA energy to just the minimum while maintaining the overall FPGA speed by reducing the supply voltage on non-critical paths. However, no existing work applied it to FPGA interconnect due to large area overhead. In this dissertation, this problem is solved by using the low-swing interconnect. When using this technique along with power-gating, a 22.3% - 56.5% energy reduction is observed. A custom low-power FPGA is fabricated and measured. When implementing a 4-bit adder, it consumes 277x lower power and 3.4x lower energy than Microsemi IGLOO, which is the most low-power FPGA in the market today.
3) Low-Power FPGA Evaluation Platform
To evaluate custom FPGAs, full tool flow and benchmark support are needed. However, existing commonly used FPGA benchmarks are either too large for ULP FPGAs or too simple to fully utilize ULP FPGA resources. Also, the existing benchmark synthesis tools either only support commercial FPGA architectures, or have crucial limitations on the syntax of input Verilog. In addition, the existing power estimation tools do not accurate for low-power FPGAs and the embedded accelerators/IPs. Solutions to those problems are addressed in this dissertation. The functionality of the custom flows has been verified using a custom low-power application suite. Compared to Microsemi IGLOO, the custom FPGA on average consumes 315x less power and 75x less energy when implementing the custom low-power FPGA application suite.
