TL;DR
The Flip FET (FFET) is a breakthrough transistor architecture that combines 3D stacking with a fully functional wafer backside, enabling signals and power to be routed on both sides of the silicon. This paper presents the first comprehensive block-level evaluation, demonstrating that FFETs can achieve a 23.3% core area reduction and a 25% frequency boost over state-of-the-art Complementary FETs (CFETs).
What Problem Does This Paper Address?
As semiconductor scaling approaches the physical limits of traditional 2D layouts, the industry is turning to 3D transistor stacking (like CFETs) and Backside Power Delivery Networks (BSPDN). However, CFETs face a major hurdle: routing congestion. Stacking transistors vertically dramatically increases pin density, making it incredibly difficult to connect everything using only the front side of the wafer.
While researchers have proposed moving signals to the backside, current architectures like CFET are asymmetric, making backside pins difficult to implement without area-intensive "bridging cells." The authors address this by evaluating the Flip FET (FFET), a symmetric architecture that naturally supports dual-sided pins and routing, potentially unlocking the next level of Power-Performance-Area (PPA) gains.
Methods
The researchers developed a full-spectrum physical implementation and PPA evaluation framework for FFETs. Key components include:
- Symmetric Cell Design: Unlike CFETs, FFETs consist of nFET and pFET pairs stacked back-to-back. This symmetry allows for identical FEOL (Front-End-of-Line) and BEOL (Back-End-of-Line) structures on both sides.
- Split Gate Technology: The separated front and backside processes enable "Split Gates," where stacked transistors can be controlled by two different signals without extra area. This is particularly beneficial for complex logic like Transmission Gates and $C^2MOS$ circuits.
- Dual-Sided Routing Algorithm: A novel flow that partitions the netlist into frontside and backside nets based on optimized pin locations, performing global and detailed routing independently on both sides.
- Power Planning with Tap Cells: To connect frontside power rails to the backside delivery network, the authors introduced "Power Tap Cells" that route VSS around VDD rails using intra-cell metals.
- Dual-Sided RC Extraction: A method to merge design exchange format (DEF) files from both sides to accurately calculate parasitics for timing analysis.
Datasets
To validate their framework, the authors used:
- Benchmark Design: A 32-bit RISC-V core.
- Technology Node: A virtual 5nm node PDK.
- Standard Cell Library: A custom-designed library featuring 3.5T (track) FFET cells compared against 4T CFET cells.
Key Insights and Ideas
- Symmetry Solves Congestion: By providing pins on both sides, FFET effectively doubles the available routing "real estate." This reduces pin density on the front side, which is the primary bottleneck for CFET scaling.
- The "Split Gate" Bonus: In sequential logic like D-Flip-Flops (DFFs), the Split Gate allows for significant area savings. For a $C^2MOS$ circuit, FFET saves 1 CPP (Contacted Poly Pitch) compared to CFET because it doesn't need extra space to route complementary clock signals to the same vertical stack.
- Cost-Friendly Scaling: The study found that FFET performance remains robust even when reducing the number of metal routing layers. A design with only 5 layers per side performed nearly as well as one with 12, offering a path to high performance with lower manufacturing complexity.
Key Results
The block-level evaluation of the RISC-V core yielded impressive results when comparing 3.5T FFET to 4T CFET:
- Area: 23.3% reduction in post-P&R core area at the same utilization.
- Performance: 25.0% higher frequency at the same utilization level (76%).
- Power: 11.9% lower power consumption at the same frequency.
- Dual-Sided Gain: By specifically optimizing for dual-sided signal routing, the researchers achieved an additional 10.6% frequency gain without any power degradation.
- Utilization: FFET reached a maximum utilization of 86%, significantly higher than the routability limits typically seen in ultra-scaled CFET designs.
Limitations and Future Work
- Power Plan Constraints: The current maximum utilization (86%) is limited by the physical placement of Power Tap Cells rather than routing congestion. Future work could optimize these power structures to push density even further.
- Cell-Specific Overheads: A few specific cell types (like AOI22/OAI22) showed slight area degradation due to the need for extra "Drain Merge" vias.
- Manufacturing Complexity: While the PPA benefits are clear, the practical manufacturing of back-to-back stacked transistors with dual-sided BEOL remains a significant challenge for high-volume production.
Summary updated