Abstract
Modern chip design is frequently bottlenecked by the lengthy synthesis process required to evaluate Power, Performance, and Area (PPA) metrics. MasterRTL introduces an innovative machine learning framework that predicts these metrics directly from Register-Transfer Level (RTL) source code, bypassing the hours-long synthesis cycle. By representing hardware as a bit-level "Simple Operator Graph" (SOG), the framework creates a universal representation that bridges the gap between high-level hardware descriptions and physical gate-level netlists. Evaluated on a diverse dataset of 90 open-source designs, MasterRTL demonstrates superior generalization and accuracy, achieving near-perfect correlation for area and timing while reducing the feedback loop to approximately 5% of traditional synthesis time.
Introduction
In the contemporary semiconductor industry, hardware engineers describe logic using Register-Transfer Level (RTL) languages like Verilog or VHDL. However, determining whether a design meets its target performance, power constraints, and area limits requires passing the code through Electronic Design Automation (EDA) synthesis tools. For complex systems-on-chip (SoCs), this "compilation" to a physical gate-level netlist can take hours or even days.
This delay creates a significant bottleneck in the iterative design process. While previous research has explored using Machine Learning (ML) to provide instant PPA feedback, these models often struggle with generalization. A model trained on CPU architectures, for instance, typically fails to accurately predict the performance of specialized AI accelerators because their high-level structures are fundamentally different. MasterRTL addresses this by moving the representation layer closer to the physical reality of the hardware.
Methodology
The Simple Operator Graph (SOG)
The core innovation of MasterRTL is the Simple Operator Graph (SOG). Instead of analyzing high-level Abstract Syntax Trees (ASTs) or complex operations (like 32-bit additions), MasterRTL decomposes the design into a bit-level graph using only five basic 1-bit operators: AND, OR, XOR, NOT, and MUX.
This bit-level representation serves as a "universal language" for hardware logic. Because SOG registers maintain a one-to-one mapping with final netlist registers, the model can perform detailed path-level analysis that mimics the behavior of the actual physical hardware, regardless of the original coding style or language (Verilog, Chisel, etc.).
Multi-Stage Timing and Power Modeling
Predicting timing is notoriously difficult due to the complexity of critical paths. MasterRTL employs a multi-stage approach:
- Path Identification: It identifies critical paths within the SOG.
- Delay Estimation: It uses Transformer or Tree-based models to estimate delays for these specific paths.
- Design-Level Calibration: To account for systematic biases—where models often become over-optimistic for large designs and over-pessimistic for small ones—a final calibration stage adjusts predictions based on the overall scale of the chip.
For power estimation, the framework breaks designs into individual modules. This granularity solves the "small data" problem by turning a single design into multiple training points. It also incorporates "toggle rates" (the frequency of signal flips), which are the primary drivers of dynamic power consumption.
Data Augmentation
To overcome the scarcity of high-quality RTL designs for training, the researchers developed a graph generative model. This model creates synthetic RTL designs, allowing the ML components to learn structural patterns even when real-world data is limited.
Experimental Setup and Datasets
The framework was validated using a comprehensive dataset of 90 open-source RTL designs, including:
- Logic Benchmarks: ISCAS’89 and ITC’99.
- Processors: VexRiscv, RISC-V cores, and Chipyard-based SoC designs (Rocket, BOOM).
- Accelerators: NVIDIA’s Deep Learning Accelerator (NVDLA).
- Open-Source Modules: Various components from OpenCores.
All designs were synthesized using a 45nm technology library to establish ground-truth labels for training and evaluation.
Key Results
MasterRTL significantly outperforms previous state-of-the-art (SOTA) solutions in both accuracy and efficiency:
- Accuracy: The model achieved a correlation ($R$) of 0.98 for Area, 0.96 for Total Negative Slack (TNS), and 0.89 for Power. Compared to previous methods, it improved the correlation for TNS by 0.33 and Power by 0.15.
- Speed: The entire estimation process—including preprocessing and inference—takes roughly 5% of the time required for full synthesis. The ML inference itself is nearly instantaneous, completed in less than 0.1 seconds.
- Generalization: Unlike previous models that were architecture-specific, MasterRTL maintained high accuracy across diverse design categories, proving that the bit-level SOG is an effective abstraction for general-purpose PPA estimation.
Conclusion and Future Work
MasterRTL provides a robust solution to the EDA bottleneck, offering designers near-instant feedback on their RTL changes. By focusing on bit-level logic rather than high-level syntax, it achieves a level of generalization previously unseen in the field.
Future research aims to bridge the "layout gap." While MasterRTL is highly effective at pre-synthesis estimation, predicting post-routing PPA—which accounts for the physical wires and congestion on the silicon—remains a challenge. Additionally, the team plans to refine the generative models to produce high-quality SOGs directly, further enhancing the framework's training capabilities.
Summary updated