Abstract: Methodologies are the key to design productivity and quality. This text is essential reading for those involved in the design of Application-Specific Instruction-set Processors (ASIPs). As silicon complexity enters the deep submicron era, traditional design paradigms face a "design discontinuity." The Mescal Methodology provides a rigorous framework to transform ASIP design from an ad-hoc process into a predictable engineering discipline, bridging the gap between the efficiency of ASICs and the flexibility of general-purpose processors.
Introduction: The Design Discontinuity
In contemporary semiconductor design, engineers are caught in a fundamental trade-off. General-Purpose Processors (GPPs) offer ease of programming and rapid time-to-market but often fall short in power efficiency and performance for specialized tasks. Conversely, Application-Specific Integrated Circuits (ASICs) provide peak performance and efficiency but suffer from astronomical design costs and zero post-fabrication flexibility.
As we transition into deep submicron manufacturing, the "design discontinuity" becomes more pronounced. The industry requires a middle ground: Application-Specific Instruction-set Processors (ASIPs). These are processors tailored for specific application domains—such as networking, graphics, or wireless communication—that maintain programmability. However, the lack of a standardized, rigorous methodology has historically made ASIP design chaotic and unpredictable. The Mescal Methodology aims to address this by providing a structured path for ASIP development.
The Mescal Methodology: A Five-Element Framework
The core of this work is a five-element framework designed to bring engineering rigor to the ASIP design cycle:
- Judicious Benchmarking: Rather than relying on generic performance scores, the methodology emphasizes benchmarks that accurately reflect the target application domain (e.g., network packet processing or video encoding).
- Architectural Space Identification: This involves defining the boundaries of the hardware design space, including instruction set architectures (ISAs), memory hierarchies, and specialized peripheral logic.
- Efficient Description and Evaluation: By utilizing Architecture Description Languages (ADLs) such as Tipi or LISA, designers can automatically generate simulators and software development tools from a single hardware description, drastically reducing manual effort.
- Comprehensive Design Space Exploration (DSE): Automated tools are employed to navigate millions of potential hardware configurations, identifying "Pareto-optimal" designs that offer the best trade-offs between speed, silicon area, and power consumption.
- Successful Deployment: To bridge the "implementation gap," the methodology provides clean programming models (such as NP-Click) that allow software developers to utilize specialized hardware features without needing intimate knowledge of the underlying microarchitecture.
Validation and Case Studies
The Mescal Methodology has been validated through several real-world domains and benchmark suites:
- Network Processing: Using NPU-specific benchmarks (IPv4 forwarding, NAT, MPLS) and the Intel IXP1200 as a case study.
- Digital Signal Processing: Evaluating performance using BDTI kernels.
- Industrial Standards: Utilizing EEMBC suites for automotive and networking applications.
Case studies on architectures like ICORE and ST200 demonstrate how instruction set tailoring—such as using CORDIC for mathematical operations or VLIW for parallelism—can significantly reduce energy consumption while maintaining high throughput.
Key Insights: Separation of Concerns and the Implementation Gap
A pivotal concept in the Mescal approach is the Separation of Concerns. By formally decoupling application requirements from architectural capabilities, designers can explore thousands of hardware variations without requiring software rewrites.
Furthermore, the methodology highlights the Implementation Gap. An ASIP's value is tied to its "Programmer's View"—the abstraction layer that enables compilers and developers to use specialized instructions effectively. Without a robust software toolchain, hardware specialization remains inaccessible and underutilized.
Conclusion and Future Outlook
While the Mescal Methodology masters the design of individual specialized processors, the industry is already moving toward the next frontier: Multi-Processor Systems-on-Chip (MPSoCs). Future research must address the challenges of inter-processor communication and the management of large, heterogeneous design teams. Despite these challenges, the shift from "logic-based" to "processor-based" design marks a fundamental evolution in how complex silicon systems are built.
Summary updated