TL;DR
OpenYieldV2 is an open-source framework that treats SRAM bitcell topology (e.g., 6T vs. 10T) as an optimizable variable rather than a fixed starting assumption. By jointly optimizing architectural choices and transistor sizing using a few-shot surrogate model, the tool achieves up to a 17.3% improvement in design trade-offs compared to traditional single-topology optimization.
What Problem Does This Paper Address?
In modern chip design, SRAM can occupy up to 70% of the die area. Designing it involves a "mixed" optimization problem: you must choose a discrete topology (the circuit structure of the bitcell) and continuous transistor sizing.
Traditionally, these are treated separately. A designer picks a topology (like the standard 6T cell) based on a rule of thumb and then uses an optimizer to find the best transistor sizes. However, topology and sizing are "non-separable"—the optimal width for a transistor in a 6T cell has little to do with its optimal width in a 10T Schmitt-trigger cell. By fixing the topology upfront, designers often miss the global optimum. Furthermore, high-fidelity simulations that include peripheral circuits and parasitics are computationally expensive, making it difficult to explore the combined design space thoroughly.
Methods
The authors introduce OpenYieldV2, which tackles these challenges through three main components:
- Dual-Topology Compiler: A fully parameterized generator that creates complete SRAM macros for both 6T and 10T topologies. Crucially, it includes all peripheral circuitry (decoders, sense amplifiers, and replica-bitline timing), ensuring that comparisons between topologies are "apples-to-apples" under realistic operating conditions. Figure 2: SRAM macro organization in OpenYieldV2.
- Few-Shot Surrogate Modeling (TabPFN): To avoid the massive cost of running thousands of SPICE simulations, the authors use TabPFN, a pretrained tabular foundation model. It acts as a "surrogate" that learns to predict Power-Performance-Area (PPA) responses from just a few hundred SPICE samples, requiring no dataset-specific hyperparameter tuning.
- Coarse-to-Refined Optimizer: The framework uses a two-stage approach. A coarse search uses evolutionary algorithms to explore discrete choices (topology, threshold-voltage assignments). Then, a refined stage uses a novel softmax-weighted gradient descent to fine-tune the continuous transistor sizes for the best candidates. Figure 1: The OpenYieldV2 pipeline.
Datasets
The experiments were conducted using the FreePDK45 (45nm) technology node. All training and evaluation data were generated internally by the tool using the Xyce parallel electronic simulator. The surrogate models were typically trained on budgets of 300 to 500 SPICE samples.
Key Insights and Ideas
- The Moving Boundary: The "best" topology isn't a fixed rule. The paper demonstrates a "topology preference boundary" that shifts based on design constraints. For example, while 6T is generally better for area, 10T becomes dominant when stringent noise margins or low-voltage operations are required.
-
Joint Optimization is Essential: Fixing a topology early discards valuable parts of the design space. Jointly searching both allows the optimizer to recover a more diverse and efficient Pareto front where 6T and 10T designs coexist. Figure 5: Joint search recovers both topologies on a single front.
-
Cross-Topology Learning: A surprising "aha moment" was that the surrogate model can reuse knowledge across topologies. Training on 6T data actually helps the model predict 10T performance more accurately, and vice versa, significantly reducing the needed simulation budget.
Key Results
- Better Trade-offs: Jointly optimizing across the topology boundary improved the hypervolume (a metric for the quality of the multi-objective Pareto front) by 17.3% over the strongest single-topology baseline.
- Surrogate Accuracy: TabPFN achieved a mean Kendall rank correlation of 0.94 with only 300 samples. In contrast, standard models like Gaussian Processes (GP) or Multi-Layer Perceptrons (MLP) struggled to reach 0.80 even with 1,000 samples.
- Refinement Gains: The local gradient-based refinement stage increased hypervolume by an additional 2.2% and reduced Inverted Generational Distance (IGD) by 55.5% compared to using global search alone.
Limitations and Future Work
The current study focuses on 6T and 10T topologies within the FreePDK45 node. The authors acknowledge that while the framework is general, the specific "decision boundaries" found might change for more advanced technology nodes (like 7nm FinFET) or for different bitcell structures (like 8T or 12T). Future work could involve expanding the library of supported topologies and testing the framework on industrial-grade FinFET PDKs.
Summary updated