Abstract
As semiconductor scaling approaches the physical limits of planar and FinFET architectures, 3D integration strategies like the Complementary FET (CFET) have emerged as the next frontier. However, CFET faces significant routing congestion due to its single-sided signal delivery. This paper explores the Flip FET (FFET), a novel 3D architecture that utilizes both sides of the silicon wafer for signal routing and power delivery. By stacking nFETs and pFETs back-to-back, FFET enables a symmetric structure with "dual-sided" signals. Our evaluation on a 32-bit RISC-V core demonstrates that a 3.5T FFET architecture achieves a 23.3% reduction in core area and a 25% increase in operating frequency compared to state-of-the-art 4T CFET designs.
1. Introduction
The transition to 3D transistor stacking is a pivotal shift in maintaining the trajectory of Moore’s Law. The Complementary FET (CFET), which stacks n-type and p-type transistors vertically, significantly reduces cell footprints. However, this density comes at a cost: routing congestion. Cramming all electrical connections onto the front side of the wafer leads to extreme congestion, often negating the area gains of the device itself. While Backside Power Delivery Networks (BSPDN) have mitigated power routing issues, signal routing remains largely confined to the front side.
The Flip FET (FFET) addresses these challenges by design. Unlike the CFET, which stacks transistors "face-up," the FFET stacks them back-to-back. This inherent symmetry allows for dual-sided signal access, effectively doubling the available routing surface and simplifying the physical implementation of complex logic.
2. The Flip FET Architecture and Implementation
The FFET architecture introduces several key innovations that differentiate it from traditional 3D stacking:
2.1 Symmetric Back-to-Back Stacking
In an FFET, the n-type logic is handled on the front side, while the p-type logic is handled on the back side. This symmetric arrangement eliminates the need for "supervias"—the deep vertical connections required in CFET to bridge stacked layers—thereby reducing internal parasitic resistance and capacitance. This symmetry also allows the cell footprint to shrink by 0.5T compared to standard CFET targets.
2.2 The Split Gate and Drain Merge
A critical feature of the FFET is the "Split Gate," which allows the top and bottom transistor gates to be controlled independently. This provides significant area savings for complex circuits like Transmission Gates and Flip-Flops. Furthermore, the "Drain Merge" naturally connects to both sides of the wafer, allowing signals to transition between the front and back sides without the need for dedicated "bridging cells" that consume valuable silicon real estate.
2.3 Physical Design Flow
To evaluate FFET at the block level, a comprehensive physical implementation framework was developed:
- Dual-Sided Signal Routing: A specialized algorithm partitions the netlist into frontside and backside nets, optimizing for the shortest path and lowest congestion.
- Power Tap Cells: Since current packaging typically delivers power from one side, special "Tap Cells" were designed to bridge frontside VSS rails to the backside power network.
- Dual-Sided RC Extraction: A novel methodology merges Design Exchange Format (DEF) files from both sides to accurately calculate the parasitic resistance and capacitance of the entire 3D structure.
3. Experimental Results
The FFET was benchmarked against a 4T CFET using a 32-bit RISC-V processor core implemented in a virtual 5nm Process Design Kit (PDK).
3.1 Power, Performance, and Area (PPA)
The 3.5T FFET demonstrated superior metrics across all categories:
- Area Efficiency: FFET achieved a 23.3% reduction in post-placement-and-routing (P&R) core area.
- Performance: At equivalent utilization rates, the FFET achieved a 25.0% higher frequency.
- Power Consumption: The architecture consumed 11.9% less power at the same utilization.
3.2 Routing Flexibility and Scaling
One of the most significant findings was the impact of dual-sided signal routing. By redistributing input pins to both sides of the wafer, the design achieved an additional 10.6% frequency boost without increasing power consumption. Furthermore, FFET proved resilient to metal layer reduction; the design maintained high routability even when the metal stack was reduced from 12 layers to 5 layers per side, offering a potential path to reduced manufacturing costs.
4. Discussion and Limitations
While FFET offers substantial PPA benefits, it introduces new manufacturing challenges. The process of "wafer flipping" and the requirement for high-precision backside processing increase manufacturing complexity compared to traditional processes. Additionally, the current reliance on Power Tap Cells to bridge the front and back sides limits the maximum achievable utilization (currently capped at approximately 86% to accommodate these taps). Finally, certain complex logic cells, such as AOI22, showed minor area inefficiencies due to the specific drain merge requirements of the FFET architecture.
5. Conclusion
The Flip FET represents a significant evolution in 3D transistor design. By leveraging both sides of the wafer for signal routing, it overcomes the congestion bottlenecks inherent in CFET architectures. The achieved 23.3% area reduction and 25% frequency boost position FFET as a compelling candidate for sub-2nm logic scaling. Future research will focus on optimizing power delivery and exploring performance at even more aggressive scaling nodes.
Summary updated