TL;DR
This paper introduces a flexible and fast power model for FPGAs, integrated into the VPR CAD tool. It accurately estimates dynamic, short-circuit, and leakage power, enabling detailed architectural trade-off analysis and the development of power-aware CAD tools.
What Problem Does This Paper Address?
Power dissipation is a critical concern in Field-Programmable Gate Arrays (FPGAs) due to their inherent flexibility, which comes at the cost of numerous pre-fabricated routing tracks and programmable switches. These components, along with the generic logic structures, contribute significantly to power consumption. While previous work has addressed low-power FPGA architectures and CAD algorithms, these solutions often focused on specific architectures or used crude power models, limiting their applicability and the accuracy of claimed power savings. This paper aims to provide a detailed and flexible power model that can be used to explore a wide range of FPGA architectures and evaluate power-aware design techniques without requiring extensive simulations.
Methods
The core of this work is a detailed power model integrated into the Versatile Place and Route (VPR) CAD tool. The model comprises two main modules:
- Activity Generation Module: This module uses the Transition Density Model to estimate switching activities within the circuit. It iteratively calculates transition densities and static probabilities from primary inputs to outputs. For sequential circuits, it extends the model to handle D flip-flops and uses a mutual probability calculation for feedback loops. To account for glitches, a "low-pass filter" is modeled at the output of each LUT, simulating the inertial delays of gates.
- Power Estimation Module: This module estimates power consumption at the transistor level. It breaks down power into three components:
- Dynamic Power: Calculated using the formula $0.5 \cdot C_y \cdot V_{supply} \cdot V_{swing} \cdot D(y) \cdot f_{clk}$, where $C_y$ is capacitance, $V_{supply}$ is supply voltage, $V_{swing}$ is swing voltage, $D(y)$ is transition density, and $f_{clk}$ is clock frequency. This is further categorized into power dissipated in routing resources, logic blocks, and the clock network.
- Short-Circuit Power: Modeled as 10% of the dynamic power, based on HSPICE simulations and FPGA datasheet parameters.
- Leakage Power: Estimated using a first-order model for sub-threshold current, considering factors like gate-source voltage, threshold voltage, and device width.
The model was calibrated using HSPICE with TSMC 0.18µm CMOS technology parameters but is designed to be general.
Datasets
Not discussed in the paper. The paper uses benchmark circuits for experiments, specifically mentioning the "20 largest Microelectronics Centre of North Carolina (MCNC) benchmark circuits."
Key Insights and Ideas
- Flexibility is Key: The power model's strength lies in its ability to adapt to various FPGA architectures by reading an architecture description file, making it a versatile tool for research and design exploration.
- Speed Matters: By avoiding time-consuming SPICE simulations or extensive circuit-level simulations, the model allows for rapid exploration of architectural trade-offs, which is crucial for large design spaces.
- Transition Density is a Good Proxy: The Transition Density Model, despite its approximations, provides good fidelity for architectural comparisons, even if absolute power values might have significant errors.
- Routing Dominates Energy: The study consistently shows that the routing fabric is the largest contributor to total energy consumption in FPGAs, highlighting the importance of efficient routing strategies.
- Short Wires = Less Energy: A clear trend emerges: shorter routing segments lead to significantly lower energy dissipation, reinforcing the value of wire locality.
- Switch Block Topology Matters: Certain switch block topologies (Imran and Disjoint) are found to be more energy-efficient than others (Wilton and Universal), offering designers a choice for optimization.
- Cluster Size Trade-offs: While larger clusters can reduce the number of clusters needed, they increase internal logic block energy. An optimal cluster size of 8-10 LUTs is identified for energy efficiency.
- LUT Size Impact: Increasing LUT size generally increases logic block energy but can decrease clock energy due to fewer flip-flops and clock branches. An optimal LUT size of 4 is suggested for energy efficiency.
- Power-Aware CAD Flow Benefits: The power-aware CAD flow consistently outperforms the baseline flow in terms of energy reduction, particularly in the routing fabric.
Key Results
- The power model achieves an average error of 4.8% when compared to HSPICE for routing segments.
- For lookup-tables, the model's predictions are within 14.5% of maximum HSPICE predictions.
- For input multiplexers, the model's predictions are scaled by 80% based on empirical results to match HSPICE, achieving better accuracy.
- For D-flip-flops, the average difference between simulation results and the model is 10.5%.
- The power-aware CAD flow generates circuits that dissipate 27% less energy in the routing fabric and 12% less energy in logic blocks on average compared to the baseline flow.
- Routing fabric accounts for 50-60% of total energy consumption, logic blocks 20-40%, and the clock network 5-40%.
- The study found that short segments are more energy-efficient than long segments.
- Disjoint and Imran switch block topologies are more energy-efficient than Wilton or Universal.
- An optimal cluster size of 8-10 and a lookup-table size of 4 were identified as most energy-efficient.
Limitations and Future Work
- Absolute Error: The authors acknowledge that while the model provides good fidelity for comparisons, there may be significant absolute errors in power estimation.
- Primary Input Assumptions: The sensitivity analysis highlights that assumptions about primary input transition density can significantly impact power evaluation, suggesting a need for more accurate input characterization.
- Routing Algorithm Impact: The choice of routing algorithm has a substantial effect on both delay and power, indicating that routing algorithm selection is a critical factor in energy optimization.
- Future Work: The paper suggests that the developed power model can be used to evaluate architectural trade-offs and the effectiveness of power-aware CAD tools. The authors' long-term goal is to further understand and investigate the impact of various architectural and CAD tool optimizations on FPGA power and energy consumption. The model is freely available for non-commercial use, encouraging further research.
Summary updated